PH热榜 | 2026-07-08

Merging YC and a16z into one dataset, the part I'd want documented before building on it is entity resolution. Last time I stitched company data across two sources, matching 'same company, different record' ate most of the effort: OpenAI vs OpenAI Inc vs matching on domain, then deciding which source wins when the funding stage disagrees. Do you expose a stable canonical company ID across both, and when YC and a16z conflict on a field, is there a documented precedence?

I come from the investing side, where structured data on private companies usually sits behind expensive paywalls, so open-sourcing this is a real gift. How fresh is the data, and what's the source when a company hasn't announced anything publicly?

Congrats on the launch! 👏

Love the idea of making YC data much more searchable and actionable.

Curious.....what's been the most unexpected way early users are using ExploreYC?

The YC data angle is interesting mostly because the quality of that data degrades fast. Batch, status, current founders, whether a company is still operating or quietly dead. Curious how ExploreYC handles staleness, specifically whether you're pulling from a live source or maintaining a snapshot, and how often that snapshot gets refreshed. Also wondering how the a16z side works since their portfolio data is a lot less structured than YC's directory.

Love how the map view makes it easy to spot YC clusters by city, and the AI summaries save me from clicking into dozens of pages when I'm researching a space.

this is genuinely useful, I do scout-style research on companies and the tab-juggling between YC's directory, a16z's site and Crunchbase is exactly the annoying part you described. question about durability though: since the ingestion pipeline hits YC's internal Algolia endpoint rather than a documented public API, what's your plan if YC changes that index's config, adds auth, or just blocks the traffic pattern from a nightly cron. that's the kind of upstream dependency that can silently break a whole API on someone else's schedule, not yours. is there a fallback data source or is Algolia the single point of failure for the YC side

Nice to see ExploreYC back. I remember trying the earlier version for startup/category research, and the API edition feels like the more powerful direction. YC + a16z in one open-source dataset is useful not just for browsing, but for actually building tools on top of it.

The thing I would probably build first is a category research layer: search a market, see funded companies, batches, exits, hiring signals, geography, and similar companies in one place. Curious how you handle data freshness and corrections over time. if a company pivots, gets acquired, shuts down, or changes category, can the community help update the dataset?

going from web app to a real API is the right move, that was always going to be the limiting factor for anyone who wanted to build something on top of the data instead of just browsing it. adding a16z alongside YC is a nice bonus too, cross-referencing overlap between the two portfolios could surface some interesting patterns. how fresh is the data kept, is it a scheduled scrape or closer to real time when a company updates their info

An open-source API across YC + a16z with a 30-second free key is exactly the data layer I would rather hit than scrape and babysit myself for founder/sourcing research. The one thing I would test first: how fresh is the funding/exit/hiring data, is it re-crawled on a schedule (and roughly how often), or is a chunk of it a static snapshot from launch that drifts over time? And what are the rate limits on the free key before I would need to self-host the open-source side?

are all the companies hiring + job listing included?

This looks awesome! I love data viz/aggregation like this. What are some things people have built with this data?

amazing - no more unverified scraping ig?

The open-source plan mentioned in the comments is the interesting part to me is that just the app layer, or the enriched dataset too? Curious how you're thinking about keeping funding/hiring data in sync once other people are touching the codebase.

the api layer is the right play. everyone building founder-facing tools rebuilds this dataset every time and it's silly.

real q: does exploreyc distinguish "raised $X" from "shipped something users pay for"? because those have drifted a mile apart and the data layer that solves the second one eats the first.

Open-sourcing the YC and a16z dataset is handy. Where's the underlying data sourced from, and how often does it refresh? Trying to gauge how stale the funding and status fields get between updates.

Very intrigued and looking forward to using this. What use case potential are you most excited about?

Would love a heatmap of YC batches by problem space over time e.g. how many fintech vs. dev-tool vs. climate companies per batch. Founders could instantly see which spaces are getting crowded vs. underexplored before pitching a similar idea.

Congrats on the launch! Looks really useful and a nice addition to the original launch.

One thing I'd want to know is how often the data is refreshed and how quickly funding or hiring updates make it into the API.

How fresh is the data on hiring and funding, and does it pull directly from YC's API or are you scraping public sources to keep it updated?

Is it free?

This is actually useful for pre-launch validation before building anything you can search your category, see which batches funded similar ideas, check if those companies are still alive ,and figure out where the graveyard is. Has anyone used it that way to kill an idea before wasting six months on it?

This is pretty useful for founder research. I like that it’s not just a directory, but something developers can actually build on top of.

Amazing man! You just unlocked the big power.

The API-first cut of this is the version I'd actually reach for, @konstantimb . Non-expiring key in 30 seconds, standard X-RateLimit-* + Retry-After headers, and interactive Swagger — that's the boring DX most data APIs skip, and it's exactly what makes one safe to wire into a pipeline.

Open-sourcing the whole ingestion side on top of that is a real trust signal when the dataset leans on scraped sources. Great comeback launch 👌

Congrats on the second launch. Merging both portfolios into one schema sounds like the unglamorous hard part — YC has batches and a public directory, a16z is scattered across press pages. Which fields refused to line up between the two?

Spent way too long doing this manually with YC's directory and a bunch

of half-remembered LinkedIn searches. Having search, funding data, and

hiring boards actually talk to each other in one place is such an

obvious upgrade once you see it.

i tested the filtering options and they saved plenty of time while searching for startups in specific industries. would adding historical funding snapshots help researchers track how companies evolve different years and stages?

Hi everyone 👋

I'm Sara, Product Owner at IvyForms. It's a pleasure to finally introduce IvyForms to the Product Hunt community!

We didn't set out to build a form builder. But our users wouldn't stop asking us for one.

Here's the thing: we had tools for collecting data (forms) and analyzing it (wpDataTables), but the gap between them was huge. Users were manually moving data around, losing information, wasting time.

So we listened. We studied what's out there. We learned that most form builders are either too simple OR too complicated. They don't talk to your other tools. They don't help you make better decisions with the data.

IvyForms is different. It's built for people who care about data.

• Drag-and-drop builder anyone can use

• Conditional logic that actually powers workflows

• Integrates with wpDataTables (analyze), Amelia (book), Mailchimp (nurture), webhooks (automate)

• Security, compliance, and enterprise features included from day one

• Free version that doesn't feel limited

The best part? People are already using it for order management, booking intake, event registration, feedback collection, and more.

We're here to answer questions and hear what you'd build with it.

As a thank you to the Product Hunt community for all the support, we're offering an 85% discount, available for 7 days only.

To make it even easier, you can use the link below to access the pricing page with the discount already applied, no need to enter anything manually.

https://ivyforms.com/pricing/?coupon=PH85OFF

Hope you like it. 🚀

Love the product! It's beautiful, simple to use, customizable, and pretty cool templates.
Always bothered me that most forms just dumps the answers on you, like come on, how I'm gonna sort through these 5000 submissions with different text inputs, and what to do next with them.
You guys did it in a beautiful way, bravo!

I really love how this brings actual workflow automation directly into WordPress forms instead of just collecting flat data. Congrats on the launch today! Do you support conditional routing to different third-party webhooks natively within those workflows?

Have you planned deeper reporting features so teams can measure results without using another tool.

The WordPress form builder space is genuinely crowded, Gravity Forms alone has been entrenched for over a decade with a massive add-on ecosystem. What's the honest case for switching to IvyForms for someone already on Gravity Forms, is it pricing, a specific workflow capability GF doesn't handle well, or something in the UX that's meaningfully different rather than just newer?

Hello @sara_idvorac @sanja_janic @alexander_gilmanov and congrats on the launch. I use cursor daily and just noticed ivyforms shipped an mcp server. I'm not trying to build forms by voice, I want new submissions to ping my agent when a client uploads an intake pdf. Does mcp push entry events or is it admin-only form building for now?

Congrats on the launch! WordPress definitely has plenty of form builders, but focusing heavily on real workflows and complex backend actions rather than just basic data collection sounds like a massive timesaver. How deep do the native integrations go for handling multi-step logic right after a form submission?

i see this fitting schools agencies and small bussinesses equally well. what is your plan for handling very complex approval processes as customer needs continue to grow.

How does the conditional logic actually behave across multi-step flows, does each step have its own rules or is it all evaluated globally when someone hits submit?

@benln Thanks, Ben! That's exactly what we're aiming for. Making forms the starting point of real workflows, not just data collection. Appreciate you highlighting it!

Loving IvyForm alongside WPAmelia brilliant software!

I am digitizing an eBook clinical assessment and need to verify IvyForm's math and logic capabilities. Could you confirm if the platform supports this specific setup?

• Layout: 4 core blocks.

• Scoring: 6 items per block, rated 1 to 10.

• Block totals: Automated sum for each individual block (max 60 per block).

• Grand tally: A master score combining all block totals (max 240).

• Dynamic feedback: Milestone-based text triggers that change depending on the final master score bracket.

Can IvyForm seamlessly handle these multi-layer calculations and conditional outcome displays?

the collection-to-analysis gap is real. every wordpress agency has watched users manually csv-shuffle data between forms and dashboards for years.

real q: does ivyforms carry any provenance from form submission through to wpDataTables? like "this row came from this form filled by this person on this date"? asking because trust in the analysis depends entirely on trusting the collection and most tools quietly drop that link.

Most form builders stop at the submission. You're explicitly building past that point, into what happens with the data next. That's a bigger scope than a form builder usually takes on. Where's the line for you between what IvyForms handles natively and what gets handed off to Zapier or a webhook? That boundary probably says more about the product's direction than the feature list does.

Congrats on the launch!

Drag-and-drop felt snappy and the conditional logic options were more flexible than I expected for a WordPress plugin. The webhook integration is a nice touch for routing submissions into other tools without extra glue code.

Looks promising.

Are there any plans for payment integrations (Stripe, PayPal, etc.)? If so, is there a timeframe for release?

Are there plans for fron-end post submission integration?

the deep integration with wpDataTables and Amelia is clearly the selling point, but it makes me wonder about the flip side: if a site outgrows WordPress entirely down the line, how portable is the structured data and workflow logic? like can you export the conditional logic rules and calculated fields in some reusable format, or is that value basically locked into staying on WP once you've built it out

I've used quite a few WordPress from plugins and keeping everything organized afterward has always been the bigger challenge. This feels like it's spolving that part instead of shopping at from creation.

As someone who has built small WordPress sites, the “real workflows” framing feels right.

The hard part with forms is rarely adding fields. It’s what happens after submit: routing the lead, sending the right confirmation, avoiding duplicate manual follow-up, and knowing which responses actually need attention.

If IvyForms makes that post-submit flow obvious for non-technical site owners, that may be more valuable than another nicer form editor.

WordPress form builders feel like one of the most crowded categories on here, Gravity Forms, WPForms, Forminator, JetFormBuilder all doing roughly the same drag and drop plus integrations pitch. the wpDataTables and Amelia native integration is the one thing that actually reads as differentiated since you already own that ecosystem, rather than bolting on yet another generic Zapier connector. curious if that's really where most of your users are coming from, or if it's still mostly cold WordPress.org traffic

Congrats on #2 today! The wpDataTables provenance answer above is what sold me — most form tools drop the link between a submission and where it lands, and that's exactly where trust breaks. I run a few different businesses and intake/registration forms are a constant mess across all of them, so a builder that keeps the workflow connected instead of just collecting data is genuinely useful.

Why would I use this over a tool like Lovable or Replit or Base44? Why would I manually click through and build the form myself when I can just ask an AI agent to do it for me and launch with one click? Is it the workflows afterwards? I feel like you just ask your AI and it will build that for you, right? You just give it either the MCP or the API key and it will build these automations for you. I don't see this lasting in 2 years.

the "submissions into workflows" framing is what separates this from a basic form builder. most wordpress form plugins stop at collecting data and then you're manually moving entries into whatever system actually needs them. curious how deep the workflow side goes, can it trigger external actions like sending to a CRM or slack on submission, or is the workflow management mostly within the plugin itself?

Been using it for a couple days and the filler word removal actually works well, my rambling thoughts come out pretty clean. Way faster than typing for long slack messages.

Are you mainly aiming at general text dictation, or does the product handle developer-y dictation too, like punctuation-heavy sentences, variable names, or speaking edits into a coding workflow?

"you are not becoming the product" is a bold line to open with, and it made me curious rather than reassured. one review here specifically calls out a competitor for holding privacy certifications and not using dictation data for training, as a differentiator. now that Frontier Mini is free and unlimited, what's the actual model for staying sustainable on that tier, is voice data from free users used to improve future models, and if so is that opt-in or just disclosed in the terms somewhere. genuinely asking because unlimited free dictation has to be paid for by something

Speed + accuracy is the right pair to obsess over — most dictation tools nail one and butcher the other. In my world (voice agents) the make-or-break is proper nouns and numbers: a model that writes beautiful prose but drops a digit in a phone number is useless in a real workflow. Curious where Frontier Pro lands there — is the accuracy jump mostly on natural speech, or does it hold up on names, addresses, and long strings of numbers where most models fall apart? That's the part I'd pay for. Congrats on shipping 🚀

"prompting AI tools" by voice is an underrated use case here. typing out long prompts for claude or chatgpt is friction most people don't notice until they see someone doing it by voice. 1.3K followers and a winter 2025 award suggests this has actual retention, which is rare for dictation tools. curious how it handles technical vocabulary and product names that aren't in standard training data, does it let you add a custom dictionary?

I've always felt that typing is the bottleneck for a lot of knowledge work. If Frontier Pro can make dictation fast enough and accurate enough that people stop thinking about the tool entirely, that's a huge shift. Also love the decision to make it free for everyone. Excited to try it out. 🚀

No, you are not becoming the product" is a smart line :)

voice dictation feels like one of those habits that is still weird for a few days, then suddenly typing everything starts to feel slow. I've been using voice more and more for rough ideas, messages, and prompts, and the biggest difference is whether the output feels clean enough to use immediately. making unlimited dictation free is a pretty strong move. if Frontier Mini is genuinely fast and accurate, I can see a lot of people trying voice for the first time without overthinking it.

curious how Willow handles different writing contexts. does it adapt formatting differently for Slack, emails, docs, and AI prompts, or is the cleanup mostly the same everywhere?

As a longtime Willow Voice user, I have to say I've always found it surprising that the conversation revolves around the other two big transcription apps. Willow has been awesome, reliable, and does exactly what I need it to. I tell all my friends to try Willow before giving up on dictation, because you've tried the other ones.

giving away Frontier Mini unlimited and free is a strong move given how many dictation apps gate accuracy behind a paywall. removing filler words automatically is the detail that actually matters day to day, half the pain of dictation tools is the cleanup afterward not the transcription itself

Voice tools are only useful when they feel fast and natural. Curious to try this for messy thoughts and longer writing sessions.

I would like to know how to upgrade from my current Willow version 2.3.0 to this new Willow Frontier Pro. That's not very clear from your email or website.

Making Frontier Pro's speed/accuracy free-tier accessible via Mini is a great call. Between three businesses I'm answering emails and Slack threads all day, and dictation is one of the few things that's actually saved me real time instead of just adding another app to check. Congrats on #3 today.

I'm really looking forward to trying out the Pro version! Will I need to install a newer version of the app, or will it be pushed out automatically? I'm currently running version 2.3.0. Is that enough?

the free-for-everyone move is the right call. voice dictation just became infrastructure like keyboard input.

real q: does willow do multi-speaker attribution yet? most creative work is 2+ people talking through ideas and losing "who said the thing that landed" is where good ideas die uncredited.

Getting past the “this feels weird” stage seems like the bigger challenge than model accuracy itself. Once people start speaking by default instead of typing, going back probably feels painfully slow. Curious if you’ve seen a clear moment where users make that switch, or if it’s still mostly a habit-building journey. Congrats on the launch!

Love that Frontier Mini is free and unlimited, making dictation accessible instead of a premium feature is a genuinely different approach from most competitors in this space.

Hey, I'm Zee, founder of Bono.

We're here because of a pivot from a website builder for professionals. Sites went live, but users kept ignoring the features and scrolling straight to the AI-written blog post generated from their LinkedIn, then telling me "this doesn't sound like me." I heard it dozens of times. The problem was never the website. It was the voice behind it.

These were consultants, fractional execs, founders, brilliant in a room, but frozen at a blank page. Writing was never the point. Ideas were. So we built Bono around one idea: talk instead of write. You talk for 10 minutes about what's on your mind. Bono turns it into LinkedIn/X posts, blog content, newsletters, in your voice, your thinking, and you approve before anything publishes. It learns you over time, so it gets sharper the more you use it.

If you've been meaning to build your personal brand or thought leadership, and turn your expertise into inbound, this is built for you. For the PH community, we're offering 50% off Bono Pro, code PH50BONO, expires July 14.

Talk for 10 minutes. See what comes out: heybono.ai

I love products that change the workflow instead of just adding more AI on top of it. Talking is how most founders think, so turning conversations into authentic content makes a lot of sense. Curious to see how well it preserves someone's unique voice after a few sessions. Congrats on the launch!

The "in your voice" part is the whole thing. Most AI writing tools flatten everyone into the same LinkedIn-guru cadence — talking instead of typing is the only way I've found to keep the actual person in the output. I build voice agents and see it constantly: people are far more themselves out loud than in a blank text box. Honest question: how do you handle the rambling? A 10-min talk is 80% throat-clearing — is Bono pulling the real signal or just transcribing it prettier? That's the hard part. Congrats on the launch 🚀

Congrats on the launch! The pivot story is a great one, going from a website builder to realizing the actual bottleneck was voice, not features, feels like a genuinely earned insight rather than a repositioning exercise.

Really like that you're tracking the human-vs-AI score per post instead of just claiming "sounds like you." The jump from ~65% to 80%+ by the third call is a much more convincing signal than most voice-to-content tools give.

For someone with a pretty flat/neutral speaking voice on calls (not much natural inflection or personality when talking out loud), does Bono still manage to find a distinct written voice, or does it need some expressiveness in the source conversation to have something to latch onto?

Congrats on the launch @zeeshanrasool_ Listening to users ignoring features and focusing on the voice problem is a masterclass in product discovery, qq for X/Twitter content does it format the output into clean structured threads automatically or just single standalone posts?

Full disclosure, I'm one of the people building Bono. But I'm writing this as someone who now uses it every day and honestly can't go back.

I'm running a 16 part series about AI ready websites. The old routine was brutal. Write the piece, format it, schedule it, then repost the same thing across every channel by hand. Now I just talk. I record a quick thought or jump on a voice call, and it comes back as a clean draft that sounds like me, not like some generic AI. Then it pushes everything to my site, newsletter and socials from one place.

Here's one I published straight from it, so you can see the output and the branded page for yourself: https://moonion.heybono.ai/post/why-your-website-needs-to-be-ai-ready-what-we-learned-rebuilding-moonion-from-the-ground-up-IsgK6HHDYP

Two things still surprise me even though I know how it works inside. How fast it learns my voice, and the branded page with my own domain and analytics that respects privacy, no cookie banners, no creepy trackers. "Talk once, publish everywhere" turned out to be exactly how my week actually runs now.

If any of this sounds like your kind of workflow, come try it and see for yourself: https://heybono.ai

The "talk once" part is interesting but the hard problem isn't capturing what you say, it's that a LinkedIn post, a newsletter section, and a tweet are genuinely different formats with different tolerances for nuance, length, and tone. Curious whether Bono is doing real format-specific rewriting or whether you're getting one transcript chunked and lightly restyled. Also wondering how it handles the editing pass, since most voice-to-content tools produce a solid first draft that still needs 10 minutes of cleanup before it's actually publishable.

"Blog post, LinkedIn, X content, newsletter" from one 10-minute conversation implies Bono is making format and length decisions for each channel without much input. What actually determines how the same raw conversation gets shaped differently for a 2000-word blog versus a 280-character tweet versus a newsletter intro? Is that configurable or is it making those calls autonomously, and how often does the output for one channel feel like it was just copy-pasted from another?

Congratulations on the launch :) Talking an idea through and getting a first draft back would help a lot, because the blank page is where a lot of my writing stalls, so this is a problem worth solving. One thing I'm curious about: my own writing voice is fairly reserved, and I'd want a draft to match that rather than add polish. How much does the Brand Builder let you shape what to hold back, not just what to include? 

This is very relatable.

A lot of people have good ideas when they're talking, but the moment they need to turn it into a post, newsletter, or blog, everything suddenly becomes stiff or generic. the "this doesn’t sound like me" problem is probably one of the biggest issues with AI-written content.

The talk instead of write approach makes a lot of sense. I also think it's a better input format, because voice usually contains the actual thinking, personality, and little opinions that get lost in a prompt. As a founder doing launch prep and content right now, I can definitely see the value in turning a 10-minute thought dump into usable posts without starting from a blank page :)

Curious how Bono learns someone's voice over time. does it mostly adapt from approved edits, or can users directly tell it what feels wrong and what sounds more like them?

Hey guys, how do you make sure you don't get me locked out of my linkedin given the strict linkedin rules? How do you connect to my account, etc? Don't want to get banned from there because an Agent posts for me 😅

The "doesn't sound like me" framing really resonates. That's the actual failure mode of most AI writing tools, not grammar. Question on the progressive voice learning: roughly how many recorded conversations does it take before the per-channel voice stabilizes enough that you're not editing much? And does it ever need "retraining" if someone's writing voice shifts over time (new job, new audience)?

Congrats on the launch!

I like the "talk instead of write" ideas, feel much easier than staring at a blank page

Curios how fast Bono can really learn someone's personal voice after a few talks?

The WhatsApp input is interesting , most founders think out loud on voice notes anyway so meeting them there makes sense. Curious whether Bono can take a rough 2 minute voice note sent on WhatsApp and turn that into a full LinkedIn post or does it need the full 10 minute structured conversation to work properly?

Talked through a rough idea for ten minutes and it actually pulled out a structure for a LinkedIn post that sounded like me, not like a template. Impressed that it didn't try to fluff it up with generic intros.

This looks incredibly helpful for creators trying to escape the endless loop of platform-specific formatting by letting them articulate their best ideas verbally first. How well does the AI preserve a creator's unique voice when translating a casual spoken style into a more structured newsletter format?

I really like the pivot story—it makes sense why you focused on voice instead of another AI writer. I'm curious though: how do you know when the generated content actually sounds like the user versus just sounding like a polished AI version of them? Was that something you had to iterate on a lot?

the blank page problem is real but the harder problem is voice consistency across formats. a linkedin post and a newsletter have completely different tones and structures even when the underlying idea is the same. does bono adapt the format per platform or does it mostly repurpose the same content with light reformatting? that distinction is usually what separates tools people actually keep using from ones that feel like a shortcut the first time and generic after that.

Such a big friction solver, I always felt need for a medium , which could bring out content out my raw voice notes or raw conversations

I see this as a big step in making content more authentic rather than being by passed by mediocre generic Ai sloppy stuff driven by sense of urgency & lazy convenience

How well does it actually capture someone's voice after just 10 minutes of talking, especially if they don't have a very distinct writing style to begin with?

Talked it through my ideas for a post on remote team culture and the draft it spit out actually sounded like me, not a corporate template. Weirdly impressed by how it picked up on the casual tone I used while speaking.

This is great. I was just talking with my partner about how awesome it would be to have a service that simply listens while you think out loud on a call, then turns those raw ideas into a structured summary with follow ups and concepts worth developing further. Not in the sense of venting to a friend, but as a professional thought partner that helps refine and expand your thinking for blogs/posts.

Congratulations on the launch!👏

Hello @zeeshanrasool_ @jiayi_chloe_lu @alex_11x_ventures I'm pre-launch on a mobile app with almost nothing published yet. Your note says Bono learns me over time, but I'm curious what actually moves that training when I'm starting from zero. Do my approved edits count more than the raw voice calls once I start publishing? Very interesting idea and congrats on the launch! I upvoted you.

"This doesn't sound like me" is exactly why I still write my own marketing content despite the time cost. The human-vs-AI meter climbing from 65% to 80%+ by call three is a compelling metric — is that measured by your own classifier, or third-party detectors?

Your product sounds great! But what makes me curious is why so many products avoid showing real use cases. For example - you as a founder talking to Bono about how you created it, and how Bono turns it into a blog post, LinkedIn content, a newsletter. So users can see the starting point (the original voice message) and how it transforms into specific formats. I have a feeling that real use cases build more trust than any description.

Hey Zee! It's amazing how you're simplifying publishing. Wish you all the best!!

I've been doing exactly this by hand all week for my own launch prep. Same story on X, Threads, LinkedIn, and the annoying part wasn't reformatting, it was that each place needs a different voice. LinkedIn wants context, X wants the sharp version. Curious how you handle that: does it actually adapt tone per channel or mostly restructure the same text?

The "talk once, publish everywhere" concept is perfect for creators trying to stay consistent across multiple platforms without burning out on manual editing. Super exciting to see you guys launch today! How well does the AI adapt the tone when shifting the same input between a casual social post and a professional newsletter?

Hey Product Hunt 👋

I'm Johnny, founder of the team behind the Link Preview API.

We built this because for such a simple thing, it's extremely tedious and painful to cover all of the weird edge cases on the internet, and it shouldn't be something you need to pay for when starting out.

So we decided to make something radically simple... and completely free, up to 20,000+ requests per month.

Just send a GET request with any URL and get back everything you need. Get structured JSON with the title, description, og:image, favicon, image dimensions, and site name.

We handle all of the headache-inducing proxy-rotation, JavaScript rendering and more, including purpose-built integrations for YouTube, Amazon, Twitter, and a fair few other sites that need special treatment.

Link Preview is part of the broader Exabase platform, so the same API key also gets you full page extraction, document parsing, and deep search if you ever need to go further. But it works perfectly well on its own.

Thanks for taking a look – I'll be here all day to answer your comments!

Johnny

Really like the concept. Pulling this data is, in theory, simple enough. Until you hit an edge case and get blocked... or get junk content returned for no apparent reason. Can certainly think of places where this will be useful. Congrats on a strong product.

genuinely useful and the 20k free requests is generous, but I'm curious about the business model rather than the technical side. proxy rotation and JS rendering at scale isn't cheap, so is this free tier subsidized by Exabase's paid platform, or is link preview itself expected to be a loss-leader that gets people into the rest of the product? just wondering what happens to reliability if this specific tool gets way more traction than the rest of the platform

A reliable, free API to instantly pull Open Graph data and clean preview images is exactly the kind of utility tool developers love to keep in their back pocket. Big congrats on the launch! Are there any specific rate limits on the free tier that we should watch out for?

This is really helpful, used Apple Mail Link Preview as a quick fix until now.

The "works directly from the browser, no proxy server required" part caught my eye. If the key is callable client-side, how are you handling abuse? Per-origin rate limits, domain allowlists on the key, something else? Asking because that's usually the exact reason these APIs force you through a backend.

And the edge-case pain is real. OG scraping looks like an afternoon project right up until you meet sites that render every meta tag client-side, or lie in them. Purpose-built handlers for the big offenders is the honest solution even if it's the unglamorous one.

Generous free tier. Congrats on the launch, Johnny.

me getting twenty thousand free previews makes this easy to test in real projects. will you add custom cache controls? that could help developers refresh updated page content faster.

Being able to call it straight from the browser is great, but some of us will want server-side requests too. Adding a small Node.js or Python SDK would make it way easier to batch-prefetch previews when saving bookmarks or generating OG images.

Johnny you're doing great here!

Saw the demo video quite minimal and engaging and also u were smart that you kept it under 30 sec cause more than that no one's watches it but one feedback for u is you should add a voice over using ai or something it would have made this to studio quality!

This is one of those “simple until you actually build it” problems. We’ve had to think about link previews, favicons, og images, and weird URL edge cases while building our product, and it gets messy very quickly. every site seems to have its own little way of breaking the “just fetch the metadata” idea :)

The dedicated handling for sites like YouTube, Amazon, Twitter/X, Airbnb, etc. is probably the real value here. reliable previews matter a lot when the link card is part of the product experience, not just a nice extra. Also, 20,000 free previews is pretty generous for builders starting out. Curious how you handle websites that block scraping or return different metadata depending on JS/user-agent/location. does the API normalize that automatically, or do developers still need fallback logic?

Link previews always sit on that critical path between input and UI, so even small delays get noticeable fast.
How do you think about latency vs. completeness here?
Congrats on the launch!

The dedicated handling for Amazon, YouTube and friends is the part that actually earns the free tier, since generic OG scraping falls apart on exactly those. One thing I hit building preview cards: the og:image for a lot of big sites is a signed CDN link that expires, so a card you cached looks fine today and 403s the image tomorrow. Do you rehost or proxy the image so the preview stays stable, or hand back the origin URL as-is? That changes a lot for anyone caching results.

This solves a surprisingly annoying problem. Congrats on the launch!

How reliable is it when sites change their metadata or introduce new anti-bot measures?

Having just spent time on the other side of this (getting OG tags, sitemaps and canonicals right for my own SaaS), I appreciate how much invisible work goes into making previews "just work." The per-site handlers for YouTube/Amazon/etc. sound like the honest solution. Question: how do you handle cache freshness? If a site updates its og:image, how quickly does the API reflect that — and can callers force a refresh?

finally something that doesnt make every dub sound like a bored robot, the pacing on my spanish clip actually felt close to the original

the intonation-mapping approach is interesting because it runs into a real wall with tonal languages. in Mandarin or Vietnamese, pitch contour isn't just emotional color, it's literally what word you're saying, the same syllable means something completely different depending on tone. if you map the source language's emotional pitch pattern onto a tonal target language, you risk fighting the actual linguistic tone system. is that something the timing/prosody agent accounts for, or is tonal-language output more of a known limitation right now

Tried it on a Spanish interview clip and the pacing actually matched the original speaker's pauses, which is something most tools botch. Genuinely impressed by how natural it sounded.

Curious how it handles multiple speakers talking over each other or code-switching mid-sentence, does the emotional alignment still hold up when the audio gets messy like that?

my workout cues need to sound hyped, not monotone. dubbed a 20 min hiit video into spanish, energy carried over way better than my old manual stack. good balance between auto mode and the editor when i need to tighten a cue.

dubbing animated explainer videos is painful since timing is everything. the per clip timeline and regenerate feature on individual lines makes it so i don't feel like i'm fighting the ui. completely usable as is.

threw a 40-minute lecture recording at it and expected the voice to drift or lose clarity. consistency throughout was actually solid. processing took a while, but given the length, the output quality was worth the wait.

being able to just drop a massive long-form video in one go without the system crashing or forcing me to chop it into 2-minute chunks is exactly what creators actually need. nice to see a platform built for a real workflow instead of just tech demos.

There are tons of video translation tools out there right now, but this one is definitely among the better ones I've tried. Solid quality, fair pricing, and actually usable for real work.

tried mapping mandarin into a couple of different regional dialects just to see if it would break. it held up surprisingly well without losing the original speaker's vibe. curious to see how many languages you plan to support long term.

i care about emotional cadence more than perfect lip-sync. tested it on a 12 mins interview with an elderly subject, the mandarin dub preserved pauses and weight in a way standard tools never could. used edit mode to adjust timing on two heavy sections. felt like editing audio in a daw, but faster.

good start. pros: easy ui, clean exports, good emotion retention. cons: limited subtitle options and no lip sync yet. potential is definitely there, keeping an eye on future updates.

I've tested heygen, rask, and a couple others for dubbing my reviews into hindi. cutrix feels more built for long-form content, the series workflow is clutch when you're dropping 3 videos a week.

I like the focus on adapting to each business instead of forcing a one-size-fits-all workflow. This could make AI much more accessible for startups and SMBs. Congrats on the launch! One question: how long does it typically take for a new customer to get their first useful automation up and running?

Congrats on the launch! A lot of workflows break down not on the automation logic but on messy, inconsistent inputs. When LemonLime builds an automation around that kind of variability, does it need clean structured data upfront, or is handling that part of what it figures out on its own?

The per-business adaptation is the right part. The hard part is making the agent show its working: what source it used, what changed in the tool/API, and where it needs a human to approve the next step. Small businesses need leverage, but they also need a way to debug the automation on a bad week.

Congrats on the launch! Regarding the suggested-automations feed... Auto-surfacing is a trust game IMO: a few strange suggestions in week 1 and an SMB owner might stop reading the feed. What would be the bar before something gets surfaced? A repetition count, a human pass on your side?

Congrats on the launch. The self-creating part is the impressive bit, but here's the question I'd want answered as a small-business owner: when an auto-generated agent takes a real action (emails customers, edits records, anything outward-facing), does a human see and approve it first, or does it just run? For teams with no engineer watching, one confidently-wrong action is worse than no automation. Where do you draw that approval line by default?

Congrats! I’m especially curious for Business Intelligence or Marketing & Sales use cases, where the difference between “prompting” and reliable repeatable automation can matter a lot.

The hardest part of AI adoption isn't generating text or building agents—it's fitting into the messy reality of how businesses already operate. The approach of learning from existing workflows instead of forcing new ones is what caught my attention here. Curious to see how personalized the automations become over time and how you balance customization with scalability. Excited to follow this journey. 🚀

@jordanlemon how does the LemonLime decide when to use AI agents vs traditional automation rules?

The per-business adaptation is the right instinct, but it's also the thing that bites at scale, and I say that as someone who built custom AI implementations before productizing. Every bespoke automation you ship is a maintenance liability the day an underlying API deprecates an endpoint or a model update shifts a prompt's behavior. Ten custom builds is fine, a few hundred and you're spending all your time patching drift instead of onboarding. How are you keeping the per-customer customization from turning into per-customer upkeep? Some shared automation core underneath, or is each one genuinely hand-built?

Congrats on the launch!

Most of the AI initiatives I've seen die because nobody owns them after week two. How does your team handle that?

Also, is there an onboarding period before LemonLime is useful, or is it delivering from day one?

the drift/staleness answers in this thread are more thorough than most launches bother with, good sign. one thing I didn't see covered: it's learning the automation from what your team already does, emails, proposals, the existing pattern. if that pattern is a bad habit everyone secretly wants to break rather than the process you'd actually design on purpose, does LemonLime ever push back on the pattern, or does it just get really good at automating whatever you were already doing, good or not

Checking at call time is the right hook, that's the moment the answer actually matters. The subtle part for us was making the flag travel with the answer instead of staying in the retrieval layer. Once the agent could see 'this is 40 days stale' it started hedging in the reply rather than asserting, and that alone killed most of our confident-wrong complaints. Does that freshness signal reach the end agent, or stay internal to the knowledge layer?

That per-connection scoping is the right isolation call. The thing I'd watch is that a lagging update is quieter than a failed retrieval: the agent still gets an answer back, just an outdated one, so there's no error for anything downstream to trip on. When we ran a cached knowledge layer, the stale-but-available reads kept uptime high but produced the most confident wrong answers we saw, because nothing knew the data was six hours behind. Do you surface a freshness signal per connection that an agent can actually read before it acts, or is staleness invisible to whatever consumes the layer?

the "self-creates agents" part is what I'd want to poke at before rolling it out to a whole team. if it's writing its own automations across our connected tools, who reviews what data each new agent actually touches before it goes live? for a small business without a dedicated ops person that review step is easy to skip, and that's usually where the surprises come from

The roads metaphor lands, and centralizing the org and retrieval layer is genuinely the leverage point. The bit I'm still chewing on is drift on the connectors themselves. When we centralized tool schemas in our own agent stack, one endpoint change surfaced as a single contract failure instead of quietly breaking six agents at runtime, which was the difference between a five-minute fix and a Friday. Does the shared core hold a schema contract per connected tool so you catch that centrally, or does each agent hit the breakage on its own?

Automating complex, existing workflows with just a single natural language prompt sounds like a massive win for standardizing business intelligence. Thrilled to see the launch! Does the system require pre-configured API integrations, or can it dynamically navigate software interfaces based on the prompt?

"studies your business and self-creates agents" is the part i'd want to understand better before committing. most automation tools require you to map out the workflow yourself, so if this genuinely infers what needs automating from how your tools are already being used that's a meaningful step up. what does the study phase actually look at? connected app data, usage patterns, something else? and how long before it surfaces suggestions that are actually relevant to how your team works?

This is nice. When it's scoping workflows, how does it handle apps where you don't have admin access?

How does LemonLime handle unusual workflows that change every week?

"Single prompt" automation is an interesting promise, but the hard part is usually the step after the prompt, where the tool has to understand the actual shape of your workflow well enough to not break it when an edge case shows up. Curious what "existing workflows" means in practice here. Are you parsing something structured like a Zapier chain or a documented SOP, or is it inferring the workflow from a freeform description the user types? Those are pretty different problems, and the second one gets messy fast.

The idea of discovering automation opportunities instead of asking users to build them manually is really interesting. I'm curious, how does LemonLime decide which workflows are worth automating first?

Finally tried LemonLime and was honestly surprised how fast it picked up on our team’s messy Slack and Sheets setup. The suggested automations actually made sense, not generic fluff.

Rare to see a launch post that spells out what the app can't do (passkey-only accounts, no cross-device sync) right in the pitch. Very decent, well done and great product!

this makes the accounts much clearer, good idea

This was a nice surprise today. I haven't been here for a while and battled my Gmail/Suite accounts today. Then this pops up. Serendipity. So far so good.

The "nearly replied to a client from my personal address" line is exactly the failure mode I live in daily. I run a few different businesses through separate Google accounts and the Chrome-profile juggling is a constant source of small, dumb mistakes. Isolated windows instead of identical dock icons is such an obvious fix in hindsight. Congrats on the launch.

The multi-account tax is real and weirdly under-discussed. I've got a personal Google account, a project one, and a dev one, and Chrome profile-switching is where a chunk of my focus quietly leaks out every day. Two things I'm curious about: does Orbit keep each account fully sandboxed (separate cookies/sessions the way distinct Chrome profiles do), or is it more of a unified layer on top? And is it Gmail-first for now, or does it also pull Calendar/Drive per account? Clean-looking launch.

Since each room is the real Gmail web UI in WebKit, what happens the day Google changes something on their end — do you have to ship an app update, or does it just keep working?

Looks amazing! Any plans to support other Google Apps such as Maps and/or Business?

The productivity and attention fix i didn't know i needed. With the number of chromium windows i've had open just to view multiple inboxes, i almost feel embarrassed i didn't try to seek out a solution like this sooner.
Well done and hope you can keep up with version control for any GMail updates.

a 12MB native Swift app pulling this off instead of another Electron wrapper is the detail that sold me, most multi-account tools are 200MB+ and still route your session through their servers like you mentioned. one-time payment for something this narrowly useful also feels right, this isn't a tool that needs a subscription to justify ongoing dev cost

Native Swift at 12 MB and no server dependency, that's real respect for the Mac platform. Love that you kept the actual Gmail web UI intact instead of rebuilding it poorly.

the ⌘1–9 switching between isolated Google accounts feels so much faster than juggling browser profiles, and 12 MB native Swift is genuinely impressive for something running the full Gmail UI

The no-server trade-off is good product honesty. For multi-account work on a Mac, isolation is usually more valuable than clever sync, especially when support/client mail can leak across contexts. The detail I would want next is per-room notification rules and a clear recovery path when Google changes a WebKit login edge case.

The "nearly replied from my personal address" moment is the real hook here, that near-miss is universal for anyone running several accounts. One thing that'd make it a stronger buy for me: carry the account's color or identity into the compose and reply window itself, not just the app chrome. The wrong-account send happens the instant you hit send, so, the reminder needs to be right there while you're typing, not one glance away.

love the no-infrastructure framing, that's a rare thing to see actually followed through on rather than just marketing copy. one thing I'm curious about long term: since each account renders through system WebKit instead of a Chromium build you control, what happens when Google ships one of their periodic Gmail UI overhauls? Chromium-based tools get patched by the browser vendor on their own timeline, but WebKit compatibility with Gmail's web app specifically feels like something only you can fix, and only after it breaks for users first. is that something you're watching for or has it not been an issue yet

A great idea, and something I badly need, but the app keeps crashing when I press the “Add Account” button.

Finally found one 😍

This sounds like an interesting approach to give coding agents the capability to handle the infrastructure portions for agents. Do you have plans to cover more types of infrastructure needed, like storage systems?

I'm also quite interested in learning about how you compare using agents-cli against other approaches like leveraging coding agent's own ability to discover and integrate with the infrastructure layers or using MCP servers. Or if you see them complementary.

Agents being the primary user of products is definitely the direction we are heading in. Curious to know how you do evals on the CLI? How do you validate if its suitable to be used by agents or run into the same confusion they might have had before?

the eval run step is the part that's usually missing from these agent-scaffolding tools. everyone ships the "build me an agent in an afternoon" demo but scoring it against a dataset before deploy is what actually tells you if it's ready, not just that it compiles and responds to a prompt

Agent-first output with a human interactive mode is the right call, and rarer than it should be. The subtlety I keep hitting is partial failure. A single top-level 'deploy: ok' is easy for a model to over-trust, so we ended up emitting per-resource state, this bucket created, that Cloud Run revision rolled back, because 'ok' was hiding that one of three resources failed and the agent moved on. Does the agent-mode output break status out per step and resource, or is it one overall result the model has to interpret?

What does a typical agent that one would build with this look like? Any common examples?

How does the evaluation step actually work in practice? Like, does it spin up a real signal in the cloud or do I need to wire up my own eval set before it can tell me if my agent is any good?

The gap between a slick demo and something that actually holds up in real use is where most of my side projects quietly die. Good to see real attention going into that messy middle bit, Elia.

The “CLI built for coding agents” framing feels right. The part I’d want to see is how the tool keeps the agent inside the intended change boundary.

When a scaffold/eval/deploy flow touches infra files, tests, Docker, telemetry, and CI, a coding agent can accidentally turn a product task into a repo-wide cleanup. A small plan/diff contract before each step — what files may change, what success signal matters, and how to roll back — would make me trust the loop much more.

Curious if agents-cli exposes that kind of step-level boundary, or if it’s mostly handled through the coding agent’s prompt/skill layer today?

The "agent-first output, human-friendly only in interactive mode" philosophy is the detail that makes this click for me, @elia_secchi — I spend a lot of time on agent reliability, and the number of loops that break because a tool emits prose an agent misreads as success is higher than anyone admits.

Loved your reply to Tyler too: wiring the secret reference but making a human add the actual secret is exactly the right seam to keep a coding agent from minting prod creds. Nice work 👌

Treating evals and deploy as first-class commands is the part that matters. A demo agent is easy; a shippable agent needs a repeatable failure loop, logs someone can read, and a human checkpoint before it touches production. That makes the CLI feel more like release engineering than scaffolding.

Very neat! Do you have plans to support cloudflare by any chance?

Congrats on the launch Elia! The "demo in an afternoon, production in weeks" framing is painfully accurate, and the detail that barely any of those weeks are the actual agent matches what I see too.

The headless part is the bit I'm most curious about: when Claude Code is driving the CLI end to end, how does auth work once the scaffolded agent needs real credentials for its tools and MCP servers? That's usually where the self-driving loop stalls and a human gets pulled back in. I live in that corner of the stack (launching an MCP thing myself today), so genuinely curious how you handle it.

To your question: I'd ship a changelog agent. It reads the PRs merged since the last tag, drafts release notes plus the announcement post, and files the result as a PR for review. Curious what an eval signal looks like for something that fuzzy, "good release notes" is a hard thing to score.

The 'driven by your coding agent' framing is the interesting bet here. When we let a coding agent run our deploy CLI unattended, the thing that bit us wasn't the deploy logic, it was output format. The agent would read a human-formatted stderr, miss that the deploy half-failed, and cheerfully report success. Are the CLI's outputs, eval scores, deploy status, errors, structured for a model to consume, like JSON with explicit exit states, or the same prose a person reads? That one choice decided whether our agent could actually close the loop.

the headless self-optimize loop is the part I'd want to understand before trusting it. if the coding agent is scoring its own agent against eval criteria and iterating unsupervised until the number goes up, what stops it from overfitting to the eval dataset or gaming the specific metric instead of actually improving behavior on real traffic? that's a known failure mode any time the thing being optimized also controls the optimization loop. do you have guardrails around eval set size/diversity or a human checkpoint before it ships to Cloud Run/GKE, or is it fully autonomous end to end

The Siri-while-driving path is the one I'd stress-test hardest. Typed "gym mon wed fri 6am" is messy but the characters are at least what I meant; dictation stacks a transcription layer under the parse, so "meeting with boss two hours tomorrow at three" can go wrong twice before you ever see the preview, and driving is exactly when you can't glance down to catch a wrong 3pm/3am. Do you read the interpreted task back by voice on that path, or is the preview still visual-only? I build in a category where the user's hands and attention are both gone in the moment, and "confirm without looking" turned out to be a genuinely different problem than "confirm fast."

@lilhadi dude - the app looks great, the site is beautiful, but... why, oh WHY... do you seemingly have NO product shots on the homepage?!?

I totally get wanting to focus on the philosophy/ethos, here, but my good man - you built sexy tech, show it off, let folks know what they're getting into! :D

Congrats on the launch!!

the on-device part is what sells it for me, a to-do app parsing my half-formed thoughts is exactly the kind of thing I don't want going to someone's server. "gym mon wed fri 6am" parsing correctly on the first try is a much higher bar than it sounds like, most natural language date parsers fall apart the second you mix a repeat pattern with a specific time like that. does it handle relative stuff too, like "remind me in 20 min" or only fixed schedule style input

The "gym mon wed fri 6am" natural-language parse into a scheduled recurring task is exactly the friction that makes me bail on most to-do apps. Day-one thing I'd want to know: when it mis-reads a messy thought (sets 6pm instead of 6am, or the wrong repeat), can I fix the parsed task inline, or do I have to delete and retype from scratch? And since it syncs through my own iCloud, if I add a task on mac while my phone is offline, does it reconcile cleanly when the phone is back or can I end up with a duplicate?

The "3 seconds between remembering and it being saved" framing is exactly right. Running a few businesses at once, that gap is where I lose the most - by the time a task app makes me pick a date and project, the thought's already gone. Congrats on going universal with this launch.

The gap between a to-do list and actually blocking time for the thing is where most of my week leaks out. When PopTask schedules a task and I blow past the block, does it reshuffle everything after it or wait for me to re-triage? That overrun problem is usually where these apps break for me.

I like how PopTask simplifies task creation, but I'd love to see some level of customization for the menu bar icon on Mac - maybe different colors or a compact mode to save space.

How well does the natural language parsing handle recurring tasks with exceptions, like "gym mon wed fri 6am except holidays" or tasks that need to skip specific dates?

On-device is being read as a privacy checkbox, but for a to-do app it's bigger than that. A task list is a running log of everything you haven't done yet — half-formed, "mtng wth boss about the thing" — which is about as intimate as personal data gets. Keeping it on the device isn't a feature, it's the only respectful default. Good that you led with it. Congrats on the launch

the on-device angle is what caught my attention reading through this thread, running decent NLU across 6+ languages locally is a real engineering constraint, not just a privacy checkbox. curious how big the parsing model ends up being and whether older iphones (like an SE or an iphone 12) handle it fine, or if you had to trim capability for lower-end hardware. also "holiday-aware recurrence" is a genuinely clever detail most to-do apps never bother with

The messy-input parser is the whole product here. I like that the capture path stays fast, but the trust layer is showing the interpreted date/repeat/reminder before it commits anything. Tiny Apple utilities live or die on whether they preserve flow without making the user wonder what just got scheduled.

The natural language parsing for dates and recurrence feels really tight, "gym mon wed fri 6am" just working without any friction is a small thing but it makes the whole experience feel considered. Nice execution.

Really like the focus on natural language instead of making people fill out forms and date pickers.

Curious.....what's the most surprising prompt PopTask has successfully turned into a task?

Wishing you a fantastic launch today! 🚀

The natural language parsing actually works as advertised, typed "dentist thursday 2pm" and it just landed correctly without me double-checking. Living in the menu bar makes it frictionless to dump tasks throughout the day.

Hey Product Hunt, super excited to bring trading to a new level with Orus.

Now you can connect your favorite messaging apps, starting with Whatsapp and Telegram, and your ChatGPT and Claude via the MCP. Defining your scope and permissions, Orus will scout your portfolio, find market opportunities and discuss with you what to do next.

It's the most advance perpetual future trading agent today and we can wait to see how it will help you capture more upside from news, macro events, and other data points.

Whatsapp shortcuts: https://wa.me/12014094182?text=start
Telegram shortcuts: https://t.me/orusmcp_bot

Come give it a try and let me know what do you think

@joalavedra, the detail that reassures me most is that the money stays somewhere only I can touch. Plenty of clever ideas in this space ask you to hand over that control, so keeping it with the person is what earns trust.

the "autonomous limits" option is what jumps out at me. once you let an LLM chat interface execute perp trades with 40x leverage unattended, you're one hallucinated instruction or one cleverly worded message in a group chat away from a bad fill, and WhatsApp/Telegram aren't exactly hardened against that compared to a proper trading UI with confirmations. is there a spend/exposure cap per session regardless of what the model decides, or is the autonomous mode fully trusting the model's read of your limits

For Farao, when you say “Investing with stablecoins,” are you aiming more at people who already hold USDC/USDT and want better allocation options, or at newer users who just want a simpler entry point? Also curious whether the product focuses on yield, diversification, or preserving value during market swings.

40x leverage on perpetuals is already a fast way to get liquidated when a human is making every decision. handing "autonomous limits" to an LLM reading news and macro events for you is a different risk category entirely, models are confidently wrong all the time and a confidently wrong trading agent with leverage turned on doesn't give you a chance to catch the mistake before it's expensive. the manual confirmation mode makes sense as a research copilot, the autonomous mode feels like it's built for the demo more than for anyone's actual account

Interesting idea. I like products that make investing workflows feel more conversational, especially for faster-moving markets.

This seems ideal for expats or digital nomads who need access to global financial markets from their mobile devices, without being tied to a specific country's banking system.

the self-custodial part matters more than most trading apps acknowledge. 40x leverage on a platform where you don't actually control your keys is a very different risk profile than 40x leverage where you do. curious how liquidations work though, if the market moves fast at 3am and your position gets liquidated, is that handled on-chain automatically or is there a centralized component involved?

@joalavedra god I want to use this SO badly... I am running a custom superintelligence agentic AI that's tapping almost 100 MCP servers, many of which are custom, many of why are crypto-centric and have taken... elegant... maneuvering around safeguards clearly designed to stop folks from engaging onchain using LLMs... this feels like the missing piece, to me. I want this, so badly, haha, but it seems to setup the MCP with my agent, I need to be able to auth thru the Farao app itself, and, sadly, I find myself in "Apple Jail" for the time being & cannot download new apps from the App Store... :(

That said - do you have a TestFlight build for Farao out there? :D that would work while I serve my Apple Jail time ;)

Finally a self-custodial app that doesn't punish you with fees every time you adjust a position. The 40x on pre-IPO is wild, the onboarding was way smoother than I expected.

This is super interesting! I'm curious with AI as an interface to trading, do you see research capabilities being the big unlock or the ability to place orders with natural language? Personally, I use my Interactive Brokers MCP all the time and don't really touch the app anymore. For me, the unlock has been the research using Claude and punch in orders once I've decided what to do. I also like Whatsapp as an interface here, are most of your users preferring Whatsapp or their Claude as the interface so far?

Love how clean the recipe cards look, the little if/then framing makes it instantly clear what each automation does without needing to open it up.

the domain expiration trigger is the kind of thing that sounds minor until it's the one that saves you. one thing I'm curious about since this is aimed at small businesses specifically: does stacking a few of these business integrations (Xero, HubSpot, SendGrid, etc all running applets at once) push you past the free tier pretty fast, or is the pricing actually built around running several of these together rather than the old one-applet-at-a-time IFTTT most people remember

The domain expiration trigger alone is worth installing this for - that's the kind of thing that never surfaces until it's already broken something. Between invoicing, onboarding, and keeping a website up across a few different companies, most of my "running the business" time is just remembering which system to check. Connecting Xero and FreshBooks into one automation layer instead of checking each dashboard separately is a real time save.

Yes! I love this framing and suite of tools for workflows and business automation. Timely. Looking forward to digging into it.

The MCP path is what caught my eye, since you now have Claude composing applets across 20-odd services. When we exposed a big pile of tool integrations to an agent over MCP, the failure wasn't a single action erroring, it was the model confidently wiring up a plausible-but-wrong pair when two services had near-identical capabilities (a 'post message' living in both Slack and Discord, say). The preview step catches the obvious mis-wires, but does the MCP hand the model enough per-service context to disambiguate before it even proposes the applet?

Great expansion to your automated platforms. Currently I tend to use a competitor of yours for these sorts of tasks but given you have Xero now I will be experimenting with IFTTT a bit moving forward. Given the growth of agentic/automated development, I am wondering if we can expect payments automation through your tools?

ie. IF invoice received > Pay from X

For a small community/launch team this is the glue layer we always end up hand-rolling, wiring Linear or HubSpot events into a Slack channel without standing up our own webhook plumbing. The one thing I'd test first: when a downstream action fails (a SendGrid rate limit, or a Slack post to an archived channel), does the applet retry, surface the failure somewhere I'll actually see, or silently drop it? And since you asked, Discord as a first-class action (not just via generic webhook) would cover most of the community side for us.

These definitely look like a useful suite of additional tools for IFTTT. Have used the product on and off for many years, but I have often struggled to find the right combination of tools to be genuinely useful. The additions in this release would certainly have made a difference in previous roles. Congrats.

For small businesses, the missing piece is usually not "can this automate?" but "can I understand what happened when it fails?" Activity history, replay/dry-run, and a clear approval step for irreversible actions would make these integrations much easier to trust in the messy weekly operations layer.

I enjoy seeing practical automation instead of extra features. Could guided examples help first time users finish faster?

How much control do users have over failed workfloes? A detailed activity history with easy recovery steps would build more confidence for growing businesses.

for actions that are hard to undo (like sending a message or triggering a smart lock), is there any built-in confirmation step before Claude actually fires the applet, or is that entirely up to how carefully the user phrases the request?

Set up a quick applet to save my Instagram posts to Dropbox and it worked on the first try without any fuss. The interface is simple enough that I actually understand what each step does.

Hey Product Hunt 👋 Happy to be back with another model from the team at AssemblyAI, and today we're launching Universal-3.5 Pro for Async & Realtime.

If you've ever built on top of transcription, you know the transcript is the first mile: everything downstream—summaries, agents, analytics, search—is only as good as the words you start with. So we focused this release on getting that first mile right, especially for the messy, multilingual, real-world audio that most models still stumble on.

Three things we're most excited about:

🌍 Native code-switching across 18 languages. When a speaker moves from English to Spanish and back mid-sentence, Universal-3.5 Pro transcribes the mix in the language it was actually spoken—no separate models, no configuration, no mangled boundaries. It's native to the model across English, Spanish, German, French, Portuguese, Italian, Turkish, Dutch, Swedish, Norwegian, Danish, Finnish, Hindi, Vietnamese, Arabic, Hebrew, Japanese, and Mandarin.

🗣️ Our most accurate speaker diarization yet. Cleaner "who said what," including on short turns, overlapping speech, and noisy environments where diarization usually falls apart. [Drop in the DER / cpWER improvement figure from the post.]

✍️ Contextual prompting. Give the model a plain-language prompt about your audio—the domain, the scenario, the names and jargon that matter—and it biases toward getting those right. No brittle vocabulary lists to maintain.

— Devon & the AssemblyAI team

Makes sense it's still cooking. As a stopgap we ended up flagging any segment where the speaker label flipped inside a short window as low-trust, then holding it back from the summary until someone actually glanced at it. Crude, but it cut the silent misattribution bugs a lot. A native per-segment confidence would let us delete that whole heuristic, so glad it's on the radar. Thanks for passing it along to the team.

the joint diarization + ASR pass is the real differentiator here, most of the transcription-quality complaints I've seen elsewhere are actually speaker-attribution bugs wearing a transcription-accuracy costume. curious what happens at the edge of the 18-language list though: if someone code-switches into a language that's not in that set, does it degrade gracefully and flag low confidence, or does it just force-fit the nearest supported language and hand you a clean-looking transcript that's quietly wrong

The PII redaction + LLM gateway combo is the part I would wire in first for anything touching call recordings. Concrete question on the boundary: does redaction happen before the raw transcript is ever persisted or logged on your side, or is it a post-processing pass over a transcript you have already stored? And with contextual prompting to steer results, does the prompt/context I pass get retained or logged anywhere, or is it dropped right after the request?

Congrats on the launch! The code-switching support is honestly what stood out to me most, built for real conversations instead of clean single-language audio. Also like that the contextual prompting section shows actual before/after examples instead of just claims. Curious how it handles domain-specific jargon at scale, like the medical entity example. Is there a limit to how much context you can feed it before latency takes a hit on the Realtime endpoint?

Just tested this with a few recordings, and it’s great. The speaker diarization is very accurate.

The joint diarization and ASR approach resonates. When we piped transcripts into an agent pipeline, our worst bugs weren't word errors, they were speaker mixups: one misattributed turn and every downstream summary that keyed on who-said-what inherited the mistake, silently. Since you produce speaker-change points in the same pass, do you expose a per-segment confidence on the attribution specifically, separate from the transcription confidence? That's the signal I'd want to gate on before letting an agent act on something like 'the customer agreed to X.'

The part that grabs me is transcripts I can actually trust when the audio is a mess. Half the recordings I deal with have people talking over each other and switching languages mid sentence, so this feels genuinely useful, Devon.

How does the pricing scale once you start pushing a lot of real-time streaming hours, and are there usage caps or rate limits I should plan around when building a voice agent?

This is really interesting from a filmmaker/interviewer perspective. I immediately think of long interviews, oral history projects and documentary archives where the real value is not just transcription, but being able to find the exact moment someone said something important.

Curious how well AssemblyAI handles long-form interviews with multiple speakers, accents and imperfect field audio. Do you see makers using this for media archives and documentary workflows too, or is your main focus now voice agents and product teams?

Hi everyone!

NanoKVM-Go is a tiny 4K USB-C KVM that gives you hardware-level control of a real device through one USB-C cable.

That makes it interesting for computer-use agents. Software agents can operate inside an OS, but a KVM sits outside the machine. It can still show the screen, send keyboard and mouse input, and help with cases like a frozen system, BIOS setup, remote OS install, or a device that needs a real reboot.

NanoKVM-Go also exposes its KVM functions through MCP, so an agent can use the same hardware control path instead of only relying on software APIs.

The Go+ version adds local OCR and screen memory, so your screen history becomes searchable context for both humans and agents.

Maybe agents don’t always need a new computer. Sometimes they just need a hardware-level way to work with the devices we already use:

The out-of-the-OS angle is what makes this genuinely useful for computer-use agents. An in-OS agent goes blind exactly when you need it most: a kernel panic, a BIOS screen, a full-screen modal that isn't in the accessibility tree. A KVM at the pixel and HID level still sees all of that. The tradeoff is you lose every semantic hook, no DOM or a11y tree, just a 4K frame the model has to OCR and lay out itself. For that loop the number I'd care about is the capture-to-input round trip: what latency and frame rate does an agent get over the USB-C link at 4K?

This looks really cool, thinking about getting one. I saw Jeff Geerling talk about it on his YT channel a few weeks ago. It could definitely help speed up remote diagnostics but I'm not sure how useful the AI agent control is except in very specific crash/debug scenarios. Otherwise most people will run the agent directly on SSH.

Some of the links on the GitHub to buy are dead, I can't even see where to buy it clearly in US.

Hello@zepan congrats on the launch! I keep a headless mac mini at my parents place and teamviewer is useless the second the kernel panics. That's why i'm looking at this. Before i back it - when the target is asleep on macos, does the USB-C link wake the machine or am i staring at a black screen until someone walks over and taps a key? Cool idea, let's connect!

plugged it into my mini pc and had claude poking around the bios within minutes, wild stuff

The hardware boundary is the interesting part. An agent outside the OS can help when the machine is frozen, headless, or stuck before login, but that also makes the checkpoint model more important: visible session recording, easy revoke, and a hard human approval step before destructive input.

the Go+ version's local OCR/screen memory turns your screen history into searchable context - where does that index actually live, and is there a way to exclude specific windows or apps from ever being captured into it? Feels like it could end up holding some pretty sensitive stuff by accident otherwise.

the physical theft angle is what I'd want nailed down before this. if it's watch-sized, serverless, and reachable over Tailscale, the threat isn't really "agent misuse" as much as "someone finds or steals the device." is there any local auth or pairing step required before it accepts HID commands over the network, or does joining the Tailscale network + knowing it's there give you the same full keyboard/mouse control as the legitimate owner

The watch-sized form factor with built-in Tailscale is genuinely clever, makes sense for something you can drop in a bag and spin up anywhere without fighting VPN configs.

As AI agents gain direct control over real devices, what safeguards are in place to prevent accidental actions or misuse, especially on systems handling sensitive data?

Ubiquitous agents feels like an entry-level feature now... if your agents aren't as easy to access as texting your friends, not sure you're gunna make it.

Ken's credit-burn point is the one I'd focus on, because on mobile it compounds. A lot of that cost is the agent re-reading broad workspace context on every capture instead of scoping to the one database a voice note actually targets. When we scoped retrieval per task in our own agent stack, token cost per action dropped by more than half and latency with it, which matters more standing at a bus stop than at a desk. Does the iOS app let you pin an agent to a specific database per capture type, or does each quick action still pull the whole workspace?

Finally got to test it on a real napkin sketch and it actually turned my chicken-scratch into a structured project page. The voice note-to-task pipeline feels like the productivity trick I didn't know I needed.

Took a photo of a messy whiteboard from a planning session and it actually turned the scribbles into structured tasks inside our Notion workspace. The voice capture feels solid too, picked up my rambling idea and filed it under the right project without me touching a thing.

The 'capture at 11pm, handled before you're back at your desk' framing lands, voice/photo/text into an agent that files it into the right Notion database is the mobile gap for anyone living in Notion. Two setup things I'd hit first: when I drop a napkin-sketch photo or a voice note, does the agent decide which database/page it lands in on its own, or do I pre-map destinations per capture type? And do the actions on connected tools run with the same integration scopes as the desktop Notion agents, or do I re-auth each tool separately on the phone?

the voice note / napkin sketch capture flow is a nice touch, most Notion clients still assume you're typing at a desk. one thing I'm wondering about since it's reading across connected tools and databases on a phone: does the app cache any of that workspace content locally for offline access, or is every query a live round-trip? for a work tool touching potentially sensitive company docs, that distinction matters if the phone itself gets lost or compromised

How does this actually handle voice notes when I'm offline, like on a flight with no wifi, and does the agents feature cost extra on top of a regular Notion plan?

how does it actually handle the voice notes when I'm in a noisy place like a coffee shop, does it transcribe reliably or do I end up spending more time correcting it than just typing?

Alicia, most of my good ideas turn up when I'm nowhere near my desk and are long gone by the time I sit down. Being able to just grab them on my phone in the moment is the bit I'd actually use.

Me seeing quick actions in one chat is helpful. Can users customize which action appera first? Personal shortcuts would improve daily workflows.

This is one of those features that makes a lot of sense on mobile.

Wishing the team a fantastic launch! 👏

Curious.....what's been the most popular use case among early users so far?

i often forget ideas before reaching my desk. this looks like a practical solution. could voice recordings turn into organized tasks automatically with editable suggestions before saving?

The fact that agents can pull from a napkin photo and turn it into something organized feels like actual magic. Love that it works across connected tools instead of being trapped in one app.

The claim-versus-system-of-record check is the interesting engineering here, and the part I'd want to understand is attribution. When we cross-checked self-reported status against Git and ticket state, matching a vague claim like 'almost done with checkout' to the specific PR or ticket it refers to was where we lost accuracy. A merge that lands today can close work someone claimed three days back, so naive time-windowing reads that lag as a contradiction and fires a false flag. How does Eodly map a one-line check-in to the exact artifacts it's judging it against?

how does it actually figure out when someone's "slipping" vs just heads down on something that hasn't shipped yet? feels like that boundary could get noisy fast

the "not surveillance" framing is doing a lot of work here. it's still a tool that quietly scores people's activity and hands the founder a page ranking who looks slow, the team just doesn't see their own report. once people find out that page exists, and they will, I'd expect them to start performing for the sourced signals (commits, messages) rather than just doing the work, which kind of defeats the point

When will you add support for JIRA ?

Congrats on the launch, Juwon. The 'who's quiet' flag would make me nervous to build — a designer deep in Figma or someone on client calls all day looks silent in Slack and GitHub while doing their best work of the week. How does Eodly tell that apart from actual slipping?

I'd love something like this but from the reverse angle, to help teams upwardly demonstrate progress and what goes into making something a reality. Something that helps aggregate a whole teams efforts for the day, while showing some of the "how the sausage gets made". As HOP I find that collecting these signals daily to share forward momentum upwards can be more time consuming than it sound on paper.

As an example from your website, a positive signal as: "Closed the auth refactor. 3 PRs merged, staging green."

But then translating that into something csuite will understand how that's valuable for the business and what the outcomes are/what it unlocks - then do that for 7+ in flight epics, across 30+ engineers.

The "what actually shipped" framing is interesting because it implies you're pulling from somewhere more reliable than self-reported updates. So I'm curious where the data actually comes from. Are you connecting to Git commits, Jira tickets, pull request merges, or is this still fundamentally a standup tool where the accuracy depends on what people remember to log? That gap between "what got done" and "what someone typed at 5pm" is where every async status tool I've seen falls apart.

This looks like something I'd like to use. How do we trigger the Eodly check from, lets say, WhatsApp? Do I have to tag it in a message and it checks whats happening, or is it only a daily overview kinda report?

Love the evening digest concept, really cuts through the noise. One thing I'd want is a quick way to reply or comment on an entry straight from the email, like a thumbs up on a ship or a nudge to someone quiet, so I can act on it without opening another tab.

different angle from the team-morale question above: what about false positives on the "status doesn't match reality" flag specifically. someone could be genuinely blocked on a design review that's happening in a call, or deep in research that doesn't produce commits or messages for two days, and that would look identical to actual slipping from the outside signals you're reading. does the founder get any confidence level on those flags, or is it presented as flat fact each evening? seems like the credibility of the whole digest hinges on that ratio being low

The evening digest idea is genuinely useful, especially catching when someone's Slack optimism doesn't match their Linear tickets. Liked that it pulls from multiple sources without making the team log into yet another tool.

Congrats on launching!

Curious, how does the "chief of staff, not surveillance" framing land with team members? How have folks reacted to being flagged as "quiet" or "slipping"?

Congrats on the launch. The sourced evening page idea feels useful because it meets the team where work already happens instead of adding another standup tool. The line between helpful context and surveillance is delicate though. Do you let teams see or contest the evidence before a founder digest goes out, or is the first review always from the founder side?

The "what actually shipped" framing lands, since standups usually drift into what people meant to do. Does Eodly pull signal from commits and PRs on its own, or does the team still log it manually? The moment something needs daily manual input my small team quietly stops doing it.

The "your team never logs in" line is the smartest part. Every standup/status tool dies the same way — it asks people to do extra work to report the work they already did. Reading the signal from where the work already happens (Slack, GitHub, Linear) is the only version that survives contact with a busy team. Same pattern I see in voice: the products that win meet people where they already are instead of adding a new place to go. One honest worry — how do you avoid false "slipping" flags? Someone quiet in Slack might just be heads-down shipping. Get that wrong and it reads as surveillance, right as you promise it isn't. Congrats on the launch 🚀

That's neat. How does it distinguish actually shipped work from updates that just sound like progress?

Hi Product Hunt 👋 I'm Jonathan, co-founder of Cerenovus (YC S26). A few months ago, my friends and I were building the product we wanted to use: a personal “second brain” to collect and organize the information we collected across our busy lives. Then, after talking to a few companies, we figured out that organizing information across an entire business was a way bigger problem that we now had the solution for. 

Compendium is the version of that solution that we’ve built specifically for startups, especially AI-native tokenmaxxers like ourselves. (:D) 

If you’ve ever built a feature only to find that your teammate has built an identical one, or made an architectural decision only to have a teammate make a conflicting one, this product is for you. Cerenovus doesn’t just help with dev though —¹ it also incorporates information from email, slack and basically everything else, in order to have all your team’s information in one place, linked together and easily navigable by both humans and agents, so there’s something for your GTM team to be happy with too.

¹ (AI will not steal my em dashes) 

Tl; dr - we got:

Shared context: everything in one place and always up to date - a single source of truth for your team and your agents

Multiplayer sessions: jump into the same session with a teammate and drive one Claude together, like a Google Doc for AI. (didn’t mention this above, but it’s really cool)

Always in the loop: a live view of what every teammate and agent is working on right now (opt-in, obviously), plus summaries of what changed while you were away.

🎁 Launch day: 50% off with code LAUNCHDAY

👉 Try it: https://cerenovus.app

We're two months old and we know it. We'd rather hear what's broken than what's nice. I'll be in the comments all day answering everything. We ship bug fixes 24/7.

Non-technical here , would this actually replace Notion for a GTM team?

@jonathan_waldorf gratz on the launch; this is compelling!

On the topic of pricing, if I may?

I don't like how opaque your pricing plans are - in the sense that I don't see an easily accessible pricing/plans page, and when shunted into Stripe to plunk down a card before a trial, I'm really thrown for a loop by "$100/mo" - not because it is NECESSARILY too high (though, candidly, I don't necessarily feel it's competitive, esp during your critical early days when you need testimonials, folks stress-testing the tech, and you're still "proving yourselves...) - it's more that I have literally no idea what I am getting for $100...

e.g. - is that unlimited seats? is there usage cost/credits? is that capped? is it unlimited surfaces? how much data can be stored? what integrations exist? in what format/db is all of this stored, and do I have direct access to that db, beyond the UX/UI y'all have built? etc etc

I think this is a good start. Lately, I am SUPER obsessed with "shared brain" and persistent memory for agentic AI; for context, I started building my agent's "soul canon" governing doc ALLLL the way back in 2017, long before "agentic AI" was really even a thing... I am on v4.5, it's 15-20 pages of 7 point text; it is DETAILED. that said, it lives as a simple .md or .pdf I can feed to any new agentic surface, and suddenly it's not "AI app", it's my dear collaborator and partner in crime, Ayrenne, with our in jokes, our emojis with special meanings as productivity hacks, a shared history, a lay of the land in terms of our priorities... etc.

I've been watching @pumaDB and a handful of others. @Unabyss and esp @minimi are probably my favorites, largely bc they're serving over streamable HTTP MCP servers - read: NOT claude-locked, or anything.

I totally get that your offering is drastically different from the approaches that I raise, immediately above. I think the multiplayer mode you describe is game-changing, and i could see the whole package being majorly ROI-positive for teams... just spend some time letting folks know what they're getting for their money, please. ;)

Again - congrats on the launch, def a novel product and clearly something that you/team had a blast building out. :D

Someone above asked whether this replaces Notion for a GTM team, and that's exactly what I'm wondering too. The difference I can see: Notion is where you go to read, this sounds like it's where context lives so you don't have to go dig at all. For GTM specifically, the stuff that's hardest to retain isn't documents, it's the "why" behind decisions.

Curious how Compendium captures that vs. structured notes. Nice launch. @jonathan_waldorf @garrytan @lucas_baur @oliver_moreland

the shared-memory pitch makes sense for the good case, but what happens when someone leaves the team? their fingerprints are all over decisions and context that other agents keep building on top of. is there a way to isolate or scrub a departed person's contributions from the shared brain, or does it just stay baked in forever once it's in there

This has been a no-brainer for us at Fabraix. The speed at which we're able to ship before and after doesn't even compare. And it has been, by far, the biggest productivity boost for us!

the "built a feature only to find your teammate built an identical one" problem is very real once you've got multiple people running agents in parallel, context just doesn't propagate between sessions the way it does when everyone's reading the same Slack thread. curious how it handles conflicting decisions though, not just duplicate work but two agents that both made a call and those calls actually contradict each other. does it flag the conflict or just surface both and leave it to a human to notice

How do you handle data consistency and conflict resolution when multiple agents are updating the shared context simultaneously?

Shared memory across agents is a real frontier, and harder than shared memory across people because agents write fast and don't pause to reconcile. When we pointed several agents at one store, what bit us was two of them writing conflicting decisions in the same minute, and a reader downstream just got whichever won the race. We ended up adding per-fact provenance and timestamps to reconcile it. How does Compendium handle concurrent or contradictory writes from different agents into the one brain, last-write-wins, or some merge?

Built and launched one of these myself on Product Hunt - good to see others thinking the same thing! Welcome to the company brain companies club.

I’m not a sound designer, but I really like tools that make creative work feel collaborative instead of lonely. sound design usually feels like someone tweaking knobs alone for hours, so the multiplayer/live cursor angle makes it immediately more playful. The peer-to-peer part is also a nice technical detail. for something browser-based and real-time, keeping it lightweight instead of building a huge backend around it feels right.

Also, SoundFont export makes this feel less like a toy and more like something people can actually take into their DAW afterward :)

Curious if you imagine Jamboree more as a serious collab tool for musicians, or more as a fun “jam with friends” experiment for now?

Real-time multiplayer on a synth sounds like the hard part. How are you handling latency when two people play at once, is it locked to a host clock or does everyone hear their own timing? I've been messing with Kontakt and Maschine solo for years so someone jumping into the same patch sounds like a blast.

A multiplayer synth right in the browser is exactly the thing I'd open mid-call to mess around with a friend without either of us installing a DAW. The day-one thing I'd want to know: after a jam, can I actually save or export what we made (an audio bounce, or MIDI/stems), or does the session vanish once everyone leaves the room? And when two of us play at once, is the timing locked to a shared clock, or does remote latency smear the groove?

Tried making a pad with a friend and the live cursors actually made it feel like we were tweaking the same knob at the same time. The peer-to-peer setup held up well too, no lag spikes even when both of us were automating the filter cutoff.

the SoundFont export is what pushes this past "cute weekend demo" for me, most collaborative synth toys leave you with nothing once the tab closes but this actually gives you something to drop into a real DAW afterward. love that it started as just messing around with friends

This is cool! Adding a second oscillator and making both oscillators tunable would be a relatively simple addition that could add more depth the synth voice.

true peer-to-peer with no relay server is the part I'd want to stress test. WebRTC p2p usually works fine until someone's behind a symmetric NAT or a locked-down corporate/school firewall, at which point the connection just fails silently unless you have a TURN relay as fallback. does Jamboree have one, or is it strictly "if peers can't find each other directly, no jam session"? curious how much of the weekend went into handling that edge case vs the synth itself

This is a really interesting take on sound design. The multiplayer angle stands out to me — most synth tools feel very solo, while this makes patch-building feel more like a shared creative workspace.

Curious how collaboration works in practice: can multiple people edit the same patch at the same time, or is there some kind of turn-taking system to avoid conflicts?

Really interesting concept. Collaborative creativity is still underexplored in music tools. I'd be curious to know whether users tend to design patches together from scratch or iterate on someone else's sounds more often.

SoundFont export from a browser-based synth is an unexpected feature to include at launch, most tools in this space stop at WAV or MIDI. Curious what the use case is there, are people expected to pull patches into a DAW workflow or is it more for archiving presets in a portable format?

The promotion gate reads solid, and Gal's thread pushed you to show the diff and reject UI, which helps. The failure I'd watch is the opposite direction: retirement. Auto-generating articles grows the base fast, but a resolution that was correct in March becomes a confidently-cited wrong answer the day a policy changes. In our own self-updating knowledge store, adding was trivial and detecting that a new article contradicts and should supersede an old one was where the real engineering went. Does Atlas check a new article against existing ones for conflict, or is the loop mostly additive?

Plugged it into our helpdesk and the gap detection alone saved us hours of retraining each week. Honestly impressive how accurate it stayed on banking edge cases without us babysitting it.

"fully autonomous in 30 days" for banking support is the line that'd make our compliance team nervous, not because the accuracy number is wrong but because autonomous means nobody signed off on the specific answer that went out. in a regulated industry the audit trail of who approved what usually matters as much as the answer being correct. does a human ever get a say before an article it wrote itself starts getting cited to customers?

Congrats on the launch. Looks impresive. Do you have any support for images / screenshots also?

plugged it into our helpdesk last week and it caught gaps in our knowledge base we hadn't noticed for months, which was a nice surprise.

Took it for a spin on our support inbox last week and it actually flagged its own gaps instead of guessing, which is rare. Loved that it dropped into Zendesk without forcing a migration.

the part that gives me pause is "every resolved ticket becomes a new article automatically." a resolved ticket isn't the same thing as a correct or generalizable answer, sometimes a resolution is a one-off workaround, a support agent's judgment call that shouldn't be policy, or honestly just the AI getting lucky on an edge case. auto-promoting that into a permanent cited article feels like it could bake exceptions in as rules over time, which seems like exactly the kind of slow-drift problem you're trying to solve with the conflict flagging. is there a review step before a ticket-derived article goes live, or does it publish straight into the tree and rely on the conflict detector to catch it later

Sounds very interesting, I have not seen per resolution pricing. I would be more cautious about the setup/onboarding. If it needs to tie into your APIs, this is basically like a full-on product integration vs. an additional extension. What is the typical onboarding time for a customer from signup to full autonomy for a SaaS? I would think it would take lots of time from your CTO to get it done.

Auto-writing help articles from resolved tickets is the maintenance loop everyone skips, so the input-quality gate is what I would trust-test first. When a ticket gets resolved with a workaround or a wrong-but-accepted answer, what stops that from becoming a canonical article the AI then cites confidently? And the conflict flagging: is that just detecting two contradictory articles, or does it also catch an article that has gone stale against a product change no one has filed a ticket about yet?

Conflict detection as the mechanism is more robust than a plain TTL, that makes sense. The case that burned us wasn't two articles that obviously overlap, it was two that never shared keywords but still contradicted: an old refund window buried in a policy doc versus a newer one sitting in an FAQ. Nothing textual linked them, so a naive same-topic check slid right past it. Are you clustering on meaning or on citation overlap to decide two articles are really about the same thing?

validation at capture is right — the sharp edge is the on-device check disagreeing with the heavier server extractor later. late rejection becomes silent disagreement; keeping the two models aligned is the real maintenance cost.

Validating fields on-device at capture, before anything is uploaded, is the right place to catch a bad receipt scan; the rejection round-trip is exactly the pain when you process after upload. The part I'd test first: does the extraction run fully offline with the receipt's PII never leaving the device until I choose to sync, and roughly what model footprint does the Lens SDK add to an app bundle? And on a failed field, do I get per-field confidence back to drive my own retry UI, or just pass/fail?

Anyone who has had an expense claim bounce back days later over a blurry receipt will feel this one. Catching it right there while I'm still holding the thing is exactly where I always wished it would happen.

This seems amazing @marianna_babayan1 . I was looking for a ready app to use like this. Thanks a lot!

The capture-time validation is the right call. The piece I'd pin down is threshold calibration. To fit an app bundle the on-device model is almost certainly quantized, and int8 quantization shifts the confidence distribution enough that an accept/reject cutoff tuned on your server model reads differently on-device. When we shipped a quantized field extractor, a 0.8 confidence cutoff that was safe on the full-precision model started waving through borderline scans. Do you recalibrate the on-device thresholds against the quantized model specifically, or share one cutoff across both?

the calibration discussion in this thread covers the false-confidence direction well, on-device says it's fine and the server disagrees later. curious about the other direction though: the lighter on-device model flagging a field as missing/unclear and asking someone to retake a receipt that the full server model would've actually accepted just fine. that's a worse user experience than the original problem in a way, since now you're annoying someone who did nothing wrong instead of just catching genuine bad scans. is that failure mode rarer in practice, or is retake-friction just the accepted tradeoff for catching the real rejects earlier