PH热榜 | 2026-05-26

Hey Product Hunt 👋

I'm Philip, co-founder and CEO of Brew. When I was leading US growth at Revolut, I kept seeing teams with world-class products fail at the last mile - actually reaching and retaining the users they'd worked so hard to acquire.

The fix was always email. It's still the highest-ROI channel for activating, nurturing, and retaining users - nothing else comes close.

But every time we tried to execute it properly, we ran into the same painful cycle. Campaigns that took weeks to build. Emails that broke in half our users' inboxes. Designs that didn't look on-brand. And when the pressure was on, they just didn't go out at all.

The problem wasn't the team. It was the tools.

Old school SaaS platforms like HubSpot, Mailchimp, and Marketo were built for companies with entire departments dedicated to running them. Someone has to analyze the data, plan the campaign, write the copy, design the emails, code the HTML, deploy it, and optimize the results.

For most teams that's just not realistic - so the software sits there underused and revenue gets left on the table.

My co-founder @thomas_park2 (former Vercel) and I built Brew to fix this.

You describe the campaign you want and Brew builds the whole thing in seconds:

• Welcome flows and drip sequences

• Newsletters and lifecycle campaigns

• Multiple on-brand variants

• Production-ready HTML that renders perfectly across Gmail, Outlook, and Apple Mail

It pulls your branding automatically from your website and Figma. What used to take even the best teams 8 days now takes one prompt.

It's like Claude Design or Lovable, but made specifically for email marketing (the emails render perfectly across inboxes because the underlying HTML is email-optimized, unlike Claude).

Watch me walk through it personally here: https://www.loom.com/share/dc325b42a4b54c50b1e5cdea3964e66a

A few things people are surprised by when they first try it:

• It builds entire multi-step sequences, not just single emails

• The brand extraction is scarily accurate

• It works natively with Viktor, OpenClaw and any AI agent

• You can push straight to Klaviyo, HubSpot, Mailchimp, or whatever ESP you're already on

This is just the beginning. We're building toward a full autopilot - you set the objective, Brew handles everything else. The goal is to become the new system of record for email marketing. One where going back to the old way is unthinkable.

Try it for free, break it, and tell us what's missing. Every piece of feedback goes straight into what we build next.

Get started free at Brew.new or reach me directly at philip@brew.new

Thanks to @jessica_w204 and @antlio for building this with me and Thomas. And to the Product Hunt community - there's no better place in the world to launch something you've worked this hard on.

Much love,
Philip 🫶

Been a happy Brew user for some time now. Incredible team and beautiful UX; this is what modern email marketing should look like.

Congrats on the launch team!

When can we expect a Brew MCP so we can automate directly from coding agents or OpenClaw? :)

I'm a happy MailerLite user, but looking at that gorgeous video I have to sign up and try the platform out! Congrats on the launch and the awesome execution :)

Congrats @philip_sorensen ! I love the demo video, inspiring to start building beautiful emails😻 How are you handling Outlook rendering under the hood? It is usually the hardest client to get right and most tools give up on it.

Looks super cool! Go go team Brew!

wow! scaling signups -> customers has been a major painpoint, and automating personalized emails seems like it would be the missing piece. LFG brew ☕️☕️ @philip_sorensen

I've been using Brew for the past few months in beta, it really does what it says on the tin. Super impressive outputs in no time!

This is so sick @philip_sorensen and team - been waiting for this!!!

Switched from Mailchimp to Brew and never been happier! Congrats on the launch guys!

Super stoked for this!

I needed to send on brand emails for one of my own side projects. Brew was a tremendous help. My favorite part is how I just gotta enter my domain and keep sending unique emails still sticking to the aesthetic, theme and taste of my brand.

Tried Brew today and it genuinely killed my last excuse for not fixing our lifecycle emails. Gave it a plain‑English prompt and got something I’d be happy to ship in one go. Super impressed with how on‑brand it comes out and how little I had to change the HTML manually.... Upvote.

Congrats on the launch, love it! A good email still goes a long way, and I think people don’t spend enough time crafting genuinely thoughtful emails anymore, instead they just blame the whole “channel” and say email is dead.

Love it! Congrats on the launch 🚀

Awesome

Finally an ESP that doesn’t feel like operating SAP through a microwave.

You can tell product taste was part of the roadmap, not just deliverability dashboards and 400 toggles.

Solo founder here, building toward an App Store launch, and lifecycle email is exactly the thing on my roadmap I keep deferring because the tooling tax is real. So the "8 days → one prompt" promise lands.

Quick q: How deep does Brew go on tone? Can it learn a brand's voice from existing copy (site, past emails), or is it pulling visual identity and writing in a competent-but-generic marketing register underneath?

The integration with tools like @Brevo is interesting. So we adopt Brew without full switching our stack. Well done.

@philip_sorensen Awesome to see native compatibility with tools like OpenClaw and Viktor right out of the gate. For an automated multi-agent workflow, can we hook Brew up to run autonomously like having an agent analyze weekly database churn metrics and prompting Brew via API to generate a custom win-back sequence?

Congrats on the launch!

I really like the positioning here. It is easy to understand, but the part that stands out to me is the last mile: not just generating nice-looking email drafts, but turning them into on-brand, production-ready campaigns and automations that can actually be sent through the tools teams already use.

As someone working around AI + productivity, I think email is one of those areas where pretty demos are easy, but reliable execution is hard. Brand consistency, lifecycle logic, ESP handoff, and rendering across inboxes are where the real value is.

Curious how you think about the rendering/testing layer over time, especially for Outlook, Gmail clipping, and dark mode. If Brew can keep that reliable while making campaign creation feel this lightweight, this could become a very useful tool for growth and lifecycle teams.

Upvoted. Excited to see where this goes.

Is there a way to easily migrate to Brew from Mailchimp or other providers?

Congrats on the launch =) One question on the brand extraction: how do you handle sites that block automated readers or render everything client-side? Curious if you've found a way around it.

congrats on the launch :) this feels like a very obvious pain point to go after. i like that it’s not just “generate an email”, but the whole messy part around brand, sequence logic, variants, and inbox-safe html.

curious how the workflow feels once brew has made the first version. do teams usually treat it as ready to send, or is there a review/tweak step before it goes live?

Been using Brew with the Customer IO integration the past two months, and it works perfectly. We send broadcast emails 5 times per week, and it's traditionally been a huge hassle to produce the content and design the emails — but it's fast (and much better) now. I especially like how we can get very differentiated versions of an email in no time. Reminds me of website building in Lovable. Highly recommend it to any marketers or founders out there!

Curious how Brew handles multiple parallel campaigns, is it one workspace per brand, or can you run separate sequences with their own logic side by side?

For context: Our team runs nurture sequences across pretty different client segments and keeping every campaign on-brand at that volume eats up real time...

Nice product btw!

The "describe a campaign in plain English, get the full thing rendered" framing is exactly what email needed — most teams aren't bottlenecked on having something to send, they're bottlenecked on the production tax of every send. I work on StoryRoute on the travel side and we see the same dynamic: people will narrate what they want a city walk to feel like, but they won't sit down and design the route step-by-step. The product that closes that gap wins. Congrats on the launch.

How complex can the automations get? Wondering if it handles conditional branches based on whether someone opened or clicked.

Hey! Amazing product.

Are there any trials available? I was looking to export the first campaign that Brew designed to me in order to evaluate the quality of the HTML.

Let me know, thanks.

👋 Hey Product Hunt,

I'm Christian, co-founder of Bond, building this with @abudi_mo and @ahx.

My last company existed to help non-technical founders build software without hiring an engineer. Thousands of them. People who had been told building wasn't for them, who sat down and shipped real, solid products anyway.

But I kept watching the same thing happen next. They'd finish the product... and then go quiet.

In an age where anyone can build, building isn't the bottleneck for growth anymore. Go-to-market is.

Doing outbound is the biggest lever. But doing outbound well today means stitching together Sales Nav, three enrichment tools, ChatGPT in ten tabs, and a copywriting tool. Then hiring a go-to-market engineer ($150k/year) to wire it all together.

Most founders can't afford that person. And honestly, they shouldn't have to.

That's why we built Bond. Great outbound campaigns that convert, run by small teams, no GTM engineer required.

Three steps:

1) Describe who you want to reach. Bond builds the full campaign plan.

2) Fine-tune and approve campaign plan. Bond executes end to end... signal triggers, enrichment, research, scoring, personalized messages.

3) Push to your sequencer. HeyReach, Instantly, or CSV. Outreach runs on autopilot.

Why it's different:

⚡ Fast. First campaign live in 15 minutes. No setup, no integrations to wire.

🧠 Sophisticated, but simple. Signal-based triggers (funding, hiring, leadership changes), 50+ data sources, every lead scored on the playbooks of top GTM teams.

🔌 One tool instead of seven. Under the hood: the 5 best email and phone finders in waterfall, real-time verification, full enrichment, research, copy. Everything top GTM teams stitch together themselves, working invisibly for you.

✍️ Personalized, not templated. Every message written from real research on that specific prospect, pushed live in one click.

The ColdIQ CEO ran a full campaign in 15 minutes with three prompts. Called it "Lovable for prospecting."

Free to start. 50% off this week for Product Hunt.

What's a signal you wish you could act on right now? A competitor raising, a target hiring their first VP of Sales, a champion switching jobs. Drop it below and I'll show you exactly how Bond runs that play.

Thanks for being here ❤️

Christian

Hey PH fam 👋

Thrilled to bring Bond to the community today. This one is personal for me.

I first crossed paths with Christian back in the peak no-code era of 2018. Watched him help thousands of founders ship real products who had been told building wasn't for them. We recently reconnected and I was struck by how locked in he is on solving the next bottleneck. Same energy, sharper focus.

He's a natural teacher, obsessive about UX, and genuinely committed to putting the power of AI in the hands of people who shouldn't need an engineer to get there. That thread runs straight through Bond.

Here's the uncomfortable truth most early teams haven't fully reckoned with:

Building is no longer the bottleneck. Go-to-market is.

The average outbound stack costs $2,000/month in tools and $150k/year in the engineer who runs them. For most early teams, that math doesn't work. That's exactly why Bond exists.

What makes it different from tools like Clay: Clay is powerful, but it's built for ops engineers who want to design workflows. Bond does that work for you. One conversation, same results, no learning curve.

The reason it works is the foundation it's built on:

→ Bond understands your company, your value props, and real buying signals -- job changes, funding rounds, hiring activity, intent data.

→ Every list is targeted because of that. Every message is relevant because of that.

→ 50+ data sources, five email and phone finders in waterfall, real-time verification -- all running underneath, invisibly. → First campaign live in 15 minutes. No setup, no integrations to wire.

Christian, Abudi, Ahmed and the team are live in the comments all day. Go check it out and show them some love. 🙏

Big congrats to the Bond team 🚀

This is exactly what I'd expect from a modern AI-first tool. Congratulations on the launch 👏

Really like the focus on non technical founders and learn teams. GTM has become way too fragmented and expensive.

I just used Bond to target every VP of Sales at a B2B SaaS company in the US who are currently hiring SDRs. The whole campaign took me under 5 minutes. No technical skills. No spreadsheets. No code. This is super cool!

How much control do you have over the ICP filters and exclusions before it starts sending anything?

15 minutes to first campaign is a bold claim for a tool that's also promising audience building, messaging, and execution. usually the speed comes at the cost of the targeting being too broad to actually convert. curious what the average reply rate looks like in practice

Bond sound like a game changer for gtm! Keep building congrats on the launch

Good tool. Not for the gtm nerds. But for someone who has never run outbound and had to rely on agencies, on heavy retainer with minimum control, this is a great alternative.

With a professional approach, there are more tools involved ;) What about checking email configuration and blacklist status? What about email warm-up? What about verifying whether a generated email actually exists? What about finding people on LinkedIn? And so on. Without all these tools, the outreach will simply end up in spam or, at best, in the Promotions tab in Gmail.

The end-to-end execution piece is what most tools in this space dodge — they hand you back to Smartlead or Instantly for the sending stack. If you're really handling warmup + sending + reply detection inline, that's a meaningful step up.

Real question: what's the data layer underneath? Clay's whole moat is the enrichment graph. Are you wrapping Apollo / ZoomInfo, or building your own?

👋 Hello Product Hunt!

Now that anyone can ship code quickly, the bottleneck has moved upstream, to the question of what gets built and why.

Rezonant sits above tools like Cursor, Claude Code and GitHub Copilot. It gives product managers a live, multiplayer workspace to turn product ideas into structured specs and tasks that AI coding agents can actually execute - grounded in the codebase, not floating around in Notion docs and Slack threads.

Capture ideas with our Chrome extension, Rezonant Alter. Hit record, point to anything on your live product, prototypes or designs, and talk through your thinking out loud, just like you would with a dev or designer. Alter captures what you said and what you pointed to, maps it to your codebase, and surfaces it as a spec or PRD ready to edit, comment on, break into tasks, and ship.

Try it for free: https://www.rezonant.app/

Can’t wait to see what you build!

@emma_burrows @vincenzo_bianco2 and team - first of all, congrats on the launch. Looks really impressive! Quick question on the Chrome extension - how does that workflow work in practice? If a PM or engineer submits feedback or feature requests through the extension, how does Rezonant turn that into something Claude Code can execute? 

Super excited to get this released and to see how teams use it!

Very excited to launch Rezonant - check out the agent document editing functionality, really great for drafting and critiquing documents (shout out to @Tiptap AI Agent)

Loved working on this! 🎉 It's really cool to be able to go back to more natural ways of communicating, bringing products to life just by chatting through ideas and features like you would with the team.

Congratulations on the launch team! Fantastic product. We’ve been using it internally for weeks and it has made a significant impact on our engineering productivity. Have shared it with several friends across startups. 💪🏻

Congrats on the launch. Most broken features I've watched ship were already broken at the PRD stage. The engineering work just compiled the misunderstanding faster. With agents grabbing tickets directly out of Rezonant, what's the QA loop before that handoff?

for me the move is the google chrome extension, then using that to screenshare and narrate directly into Rezonant to get that turned into tickets and specs. Loving it so far, like a vibe-product platform :)

Love this! The flow mode to PRD is great. Also enjoying sending straight to coding agents for simple tasks.

@abi_church Congratulations on the launch

It's a pleasure to work together and see your launch! I wish you lots of success moving forward.

Congrats on the launch @emma_burrows !

So this is automatic PRD creation, right? Do you also have a way to connect this with code agents like Codex or Claude code to implement directly on codebase too?

Congrats on the launch! Very timely idea. More teams are building with AI coding agents now, but the messy part is still aligning product context, specs, and execution. Rezonant seems to tackle exactly that gap. Love the workflow-first approach.

Is this for beginner teams that don’t have business analysts? What you described is usually done by a business analyst. Also, now everyone uses Claude Code, Claude Design, Figma, and other tools, so often you need not just a comment but specifically a comment in, for example, Figma, so that Claude Code can later understand it and apply the change. So at the current stage of AI development, I can’t imagine how to do this without a human.

Congrats on the launch!

ChatGPT-vs-Rezonant comparison on the landing page is the best pitch I've seen this week, where Rezonant pulls the actual src/services/integrations/ path instead of giving a generic playbook.

We're building in the SDLC execution space at Revolte (the agent runs from spec through deploy), so the spec-quality problem hits us directly downstream. Curious how you're handling the codebase index — persistent semantic index per repo, or retrieval at refinement time? We went persistent and the freshness problem is harder than I expected.

"Turn messy ideas into structured specs" is a deceptively hard problem because the messiness is where domain knowledge lives — strip too much of it out and the spec becomes generic; leave it in and the agent gets confused. I work on ModeLoop in financial modeling and the same pattern shows up there: the difference between a model an agent can build and one a human can defend is usually in the assumptions, not the cells. Curious how you've handled assumption capture in Rezonant.

We're drowning in 'quick ideas'/proof of concepts/unclear specs - great to have something that can streamline and get the right context all together!

Spec-to-code via voice is a workflow I didn't know I wanted until now. Product teams burn so much time translating verbal ideas into structured tasks. We've been building in the ops-heavy SaaS space, and this kind of frictionless spec creation is something PMs ask about constantly. How does Rezonant handle ambiguity in voice input, does it ask clarifying questions?

Grounding specs in the actual codebase is the real differentiator here, but codebases move every day — does the grounding re-run, or is a spec a point-in-time snapshot that quietly goes stale the moment someone merges a refactor? The failure mode I'd watch for is a confident spec citing a service or path that got renamed last week.

Hey Product Hunt! 👋

Quakpit started as a side experiment, I wanted to build my first macOS app and see if I could actually ship something.

So I built a tiny app that sends a duck-piloted plane flying across your screen when a meeting is about to start, with a banner showing how many minutes you have left. I posted one tweet about it with zero expectations.

It blew up.

Hundreds of people asked for the same thing: more animals, custom banners, sounds, calendar integrations. Turns out nobody actually likes boring notification popups.

So I built it properly. Here's what Quakpit does today:
🦆 Connects to your calendar and triggers a flying animation before your meetings
✈️ A plane crosses your screen with a banner showing your meeting name + time left
🦕 Choose your pilot — duck, dinosaur, pigeon, capybara, or dog
🆓 Completely free and open source
⭐ One-time $4.99 premium for custom banners, sounds, animals and more

It's built with love. And the open source part is real, the full free experience is on GitHub.

Would love to hear which animal you'd pick as your co-pilot 👇

It looks cute! Will the notification be seen to others if I'd sharing the screen during the video call?

Okay this is legit awesome. Going to try to download and install now!

Will provide feedback if I have any :).

this is the first calendar tool i've seen in years where the feature set made me actually laugh. open source too. what made you pick the plane mechanic specifically over something like a walking animation or a popup

Okay, this is actually a fun idea 😄 Finally a meeting reminder that doesn’t feel depressing. A capybara flying across my screen would definitely get my attention.

The category of "reminders that don't feel like nagging" is underrated — the second a notification starts feeling like work, people swipe it away and the whole system loses its job. A duck towing a banner is doing real product work there. I build StoryRoute on the travel side and we've found the same lesson: the delight isn't decoration, it's the only reason the user gives the tool another chance after the first miss. Congrats on shipping.

so cool! hit the upvote button 3 times!

Congrats with the launch! Capybara pilot is GOAT 🫶

Like the style! So cute =)

I mean...I just have to clap. Always love these creative ideas coming to life and productized. :D

capybara, no contest. if an unbothered little guy is calmly flying across my screen telling me i’m late, i’ll probably listen.

love the mix of useful + ridiculous here. boring meeting popups are way too easy to mentally mute.

Saw this on Twitter the other day. Super fun.
These are the kinds of apps I want to see here more often :)

First feedback: Why not access the local apple calendar directly? So I don't have to connect to iCloud / Google?

This is so unnecessary and fun :)

Genuinely the most charming product on PH today.
Somehow a duck towing a banner is harder to ignore than the macOS notification I dismiss reflexively a hundred times a day.
Honest question: is the capybara winning in your usage data? I need to know if my taste matches the median user.

Your logo made a difference, so I clicked! :)

This is such a delightful app. I'll get excited for every meeting if this is how I'm reminded of it.

I just sent this yesterday as a reference to a friend for great solutions for people who need novelty to get things done. Love this, Tom.

P.S. Also loving the banner going through the Ooble Studio website.

This is fantastic! I love joining meetings now

Love this. Such a small idea but way more memorable than another boring calendar notification. The flying plane + animal pilots makes it really fun. Clean execution and open source is a big plus.

Corgi is definitely my co-pilot 👨‍✈️🐶

Hey Product Hunt 👋

Thrilled to introduce Parrot, Ringg’s speech-to-text model built for production-grade voice agents.

Most STT models do well on clean audio. Voice agents don’t get clean audio. They deal with compressed phone calls, Hindi-English code-switching, Indian accents, background noise, and conversations where one misheard word can break the next action.

What makes it different:

🦜 Built for real world calls
🦜 Low latency inference for smoother voice agent conversations
🦜 Hindi validation and normalization for cleaner downstream workflows
🦜 Strong Normalised WER performance on open-source Hindi benchmarks

For teams building voice agents, Parrot helps turn messy speech into cleaner transcripts that LLMs can actually use.

Try it out and let us know what you're building with it!

Best for voice AI use case!!

Haha, how can something be this useful and this scary simultaneously!? As someone with a name most humans can't spell right, I look forward to the day when this is no longer an issue.

Try this out with easy to integrate package https://www.ringg.ai/dashboard/stt

This looks really solid 🔥
Curious about latency and how it performs in noisy real-world calls compared to Whisper.

Congrats on the launch! Building voice sessions into a couples app right now (currently on Deepgram for streaming transcription), so the "voice agents don't get clean audio" framing really lands...clean-audio benchmarks oversell every STT model until you hit a real room. One thing I've run into that I'd love your take on: the hardest case isn't accent or noise, it's two people talking, overlapping speech, interruptions, one person finishing the other's sentence. Most STT degrades badly there. Is Parrot tuned mainly for the single-caller voice-agent case (one human, one agent), or does it hold up on genuine multi-speaker conversations? Curious whether that's a roadmap item or a deliberate scope line.

Just tried this out, amazing speed and accuracy. Great work!

Our AI has a voice mode. We use ChatGPT (before that, we used speech-to-text recognition and then text-to-speech). How is your service better? Is it only better Hindi recognition, or is there something else?

Building a dedicated validation layer for Hindi downstream workflows is clever. Most generic STT APIs fall apart on code-switching and regional accents. We've hit similar walls where raw transcripts were too noisy for reliable intent parsing in production pipelines. How do you handle Hinglish code-switching, and what's the P95 latency on a 10-second audio chunk?

Congratulation on the launch! Btw, when I mix English with Hindi, I observed its little biased towards transcribing English in Hindi (using Devnagri glyph). Latency is impressive

Production-grade STT that holds up on noisy, code-switched real-world audio is harder than the demos make it look. Hindi-heavy + noisy conversations is exactly the unglamorous evaluation set that exposes most general-purpose STT models. I run a finance podcast (ModeLoop Podcast) and the transcript-quality drop between studio-clean and live-recorded episodes is enormous; tooling that closes that gap meaningfully on top of low latency is genuinely useful. Curious whether you're benchmarking against WER on standard sets or against task accuracy on downstream agent workflows.

The Hindi-English code-mixing capability is the genuinely hard part here. Most STT models either treat it as two separate language passes or degrade significantly at switch points mid-sentence. How is Parrot handling segmentation at the language boundary? Specifically, when a speaker switches mid-phrase rather than mid-sentence, does the model maintain a single continuous transcript or does it stitch segments, and how does that affect downstream NLU latency?

Hey Product Hunt 👋

I'm Sergei Sherman, CEO of @Avaturn.

Today we're releasing AVTR-1 — an open-weights real-time AI avatar model that sets a new state of the art on key benchmarks.

If you're building anything with real-time AI avatars, AVTR-1 is for you.

✍️ Here's what makes AVTR-1 different:

1. The whole face is generated. Not just the lips swapped onto a pre-recorded clip. Every pixel of the avatar's face, top of the head to the chin, is generated in real time, frame by frame.

2. Native duplex — the avatar actively listens. The model is generating all the time, whether the avatar is speaking or listening. Just like a human on a call, the avatar's face responds to your words and your tone in real time. The brow lifts at word three because you sounded surprised at word three, not after the sentence ends.

For three years, "real-time avatars" have meant pre-recorded video with a generated mouth pasted on top. We threw out the recording.

🎯 Why you want AVTR-1:

• Open weights. Free for personal, research, and any commercial use under $10M in annual revenue. Commercial licensing above that, through us.

• Sub-200ms end-to-end on one A100 or 4060. Runs on youd device in a data center, in the cloud.

• Avaturn Streamer included — the open infrastructure layer for real-time avatars. Accepts AVTR-1 or any other open-weight real-time video model as a drop-in. Plug in your video model on one side, your conversation backend on the other.

• Reference avatars out of the box. Model cards, license-cleared, deployable today.

• Launch-partner examples in the repo with Cartesia and Pipecat on day one.

🏗️ One thing we're explicitly NOT launching — and want the industry to build with us:

A public, vendor-neutral leaderboard for real-time AI avatars. The category needs a transparent scoreboard, one the ecosystem runs together. Clear, public competition is the only way improvement happens fast.
We're inviting every other vendor, every open-source contributor, every researcher to help us build it.

🎉 Everything is live today:

• Code, inference, evaluation: github.com/avaturn-live/avtr-1

• Download Model Weights huggingface.co/avaturn-live/avtr-1

• Technical report, full paper, reproducible benchmarks: avtr-1.avaturn.live

• Hosted demo: avaturn.live

Real-time generated video is the next frontier. Every previous wave — text, then real-time audio — produced an open layer the category built on. We're shipping that layer today: model and orchestration both.

Drop questions, feedback, or what you're building below — I'll be here all day 🚀

— Sergei

the active listening part is what separates this from every other avatar tool. current ones just talk at you with dead eyes while you wait. if the expression matching actually works in real time this changes how you build AI sales and onboarding flows

Tried the demo before launch — the lip sync is noticeably better than what I’ve seen with other generative avatars. But @sergei_sherman walked me through something deeper: active listening coupled with empathetic response.

You know how you can kind of say anything to most AI avatars (e.g. “my mom died”, or "omg there's a murderer outside my window!") and they'll just blink, cycle their idle loop, nod and say something like “oh, that’s nice to hear.” These bots are just mouths on a timer with zero semantic read.

AVTR-1 generates every pixel of the face in real time, frame by frame. When meaning shifts in what you’re saying, the expression shifts to match — e.g. brow lifting at word three because the content warranted it, not just because the sentence ended.

For developers: there’s no Pipecat equivalent for video agents right now. @Avaturn Live is shipping the full stack — model weights, streamer, sync layer, reference avatars. Bring your own GPU, and you're ready in 15 minutes. Open weights is a big deal, and it's all free if your business is under $10M ARR.

Hey guys! We are so excited to show you our new model: avatars became even more realistic and reactive. The awesome thing is that active listening is now at another level: avatars are reacting to your speech like a real person. If you are not a technical person like me, you can simply go to our website and talk to our avatars to see how cool that is! If you are a developer yourself, check out our github: we opened our model!

the 9x faster claim is interesting but faster than what exactly. the baseline matters a lot here. sub-300ms felt like magic two years ago and now it's table stakes for anything calling itself real-time. curious what the actual latency numbers look like end to end, not just the generation side

How is it different from Tavus or HeyGen?

What GPU is needed?

Full-duplex with "minimal latency" and every-frame-generated is the hard combo — what's the actual end-to-end latency, and on what GPU? Open weights only matter if a small team can self-host without a rack of H100s, so the number that decides adoption is "real-time on a single 4090" versus "real-time only on datacenter silicon." Which is it today?

Congrats on this amazing release! I’ve been in the real time Avatar business for the past 20 years. My first real time Avatar engine, a desktop windows app called Virtual Assistant Denise was released in 2008, and finding out you’re open sourcing your framework is just a unbelievable good news taking in consideration the high quality and competitiveness of your engine. You may be right now disrupting a very competitive market, where most big players will have to review their pricing. I have great respect for all those current Avatar companies as I’ve worked with most of them and know their very skilled people and their effort to stay competitive, and I do understand their efforts and investments to stay alive. But on the other hand, they need to go back in time and observe what companies like Unreal did to survive when Unity came up and disrupted the game market. Thanks to that, today we have small group of people releasing amazing games! Congrats again on this intrepid move, as developers can now focus on the creation of the final product, and not on finding ways to pay for subscriptions. TTS, STT, LLMs, memory and database frameworks were until a few time ago in the hands of a few companies and today they became commodities. Your decision to open source this engine is one big step to democratize this important piece of software to build human machine interactive product possible for everyone.

Open-sourcing the weights for a real-time avatar model is a much bigger deal than the headline suggests. The closed-stack incumbents in this space charge per-minute and effectively gate experimentation to whoever can afford to burn API credits playing with use cases. I make finance educational content on Mod3Loop (YouTube) and the choice between "talking head on camera" and "avatar reading a script" has been completely blocked on per-minute cost economics for indie creators — free weights changes that math entirely. Full-duplex listening while speaking is the unsexy part that actually makes these feel like presence rather than playback.

Really great to see truly open models for realtime video (and a nice technical overview post). Congratulations on the launch, team!

Love this! I am working on a project to enable voice as primary interface for agents doing real work would love to integrate with AVATURN! Tried using with Tavus before but too pricey, how can we scale this to multiple users any tips on the same? Love the project and initiative, congrats on the launch!

Tried the demo, looks impressive. Why have you decided to open source your model?

Great work on this launch! Running inside the apps people already use instead of asking them to change their workflow is exactly the right call

The interesting bit here is the gap between “how I talk while thinking” and “how I want the final message to read.” A lot of voice tools preserve the first too literally, which makes the output fast but still not quite sendable.

I’d be curious how Willow separates transient spoken scaffolding — false starts, caveats, messy ordering — from durable voice traits like sentence rhythm, level of warmth, and how direct someone tends to be. That feels especially important inside Slack/email, where the same person may want a very different register depending on recipient and risk.

The biggest issue for me has always been that I don’t speak in perfectly structured sentences, especially when writing emails, messages, or prompts. So the idea of turning a rough spoken thought into a clean message in your own voice feels much more useful than traditional dictation.

The context-aware replies are especially interesting. If it can understand the thread and draft something that actually fits the situation, that could save a lot of time. A few questions: how much does Willow learn from someone’s writing style over time, and can users control or reset that personalization? Also, how does pricing work after the Product Hunt discount?

Really curious to see where this goes. Good luck with the launch!

honest question after 6 launches: what's the retention curve actually look like for scribe specifically versus the base dictation product. dictation has obvious daily utility because it replaces typing. scribe feels more like something people try for a week and then forget exists unless it's deeply in the habit loop. how are you solving for that

I’ve been a heavy user of Claude and GPT for a long time, especially for writing emails and posts. They’ve been a core part of my workflow.

Recently, I started using Willow Scribe (Ive used regular Willow dictation for about a year), and I absolutely love it. I’m using it to draft emails, create LinkedIn posts, refine prompts, and even support a white paper I’m currently writing. It feels much more natural and fluid, especially compared to traditional dictation tools that capture every false start and half-formed thought.

Willow has genuinely improved how I think and write in real time. I couldn’t recommend it more.

The "removes filler words" line is the actual unlock here — most dictation tools dump a raw transcript that still takes 30 seconds to clean up, which defeats the whole point.

Quick question: can I toggle the cleanup off for casual Slack messages where I want my real voice? Sometimes "um, actually" is the message.

The highlight-anything-and-rewrite-in-place feature is the one I'd use most, and also the one I'd poke at first — does that work through the macOS accessibility layer in any text field, or only inside the named apps (email, Slack, iMessage, Docs)? The "in place" part is what separates this from copy-pasting into a chatbot, so how far the surface actually reaches matters.

Congrats on the launch!

Hey Product Hunt 👋

I'm Dr. Rishi Thussu from SelectPrism. I spent years at Monster watching millions of applications pile up and go nowhere. Then at Findem, I was building AI tools for talent intelligence — and seeing what was actually possible when you applied real machine learning to this problem.

Both taught me the same thing: the problem was never finding applicants. It was what happened after.

SelectPrism is an AI interviewer that runs your entire first round, from screening to scoring, so your team shows up only for the conversations that matter. It plugs directly into your existing ATS, so every candidate, report, and hiring decision stays in one connected flow. 

We built this for Talent Acquisition (TA) leaders who are great at their jobs but spend most of their week on work that no longer requires a human.

If you're losing good candidates to slow pipelines, this is for you.

The problem we kept running into was simple:

The best candidates don't wait two weeks for a callback. By the time most TA teams finish screening and scheduling, the people worth hiring have accepted other offers.

That's why SelectPrism is built around one complete flow: screen → interview → shortlist.

Here's what that looks like in practice:

📋 Drop in a job description. SelectPrism parses it, maps it to a skill graph trained on 60M+ profiles, and ranks your applicants by actual fit, not keyword matches.

📩 Interview invites go out within the hour. No scheduling, no back-and-forth. Candidates get a link and show up when they're ready.

🎥 The AI conducts the interview. Live video, adaptive questions that adjust in real time, voice-based pre-screening, and coding assessments where the role needs them.

🔒Holistic anti-cheating checks. 25+ integrity checks across browser, device, and video — including gaze detection, lip sync analysis, AI voice agent detection, candidate ID verification, and real-time fraud flagging. 

📊 Your recruiter gets a decision-ready report. Fit score, strengths, gaps, full transcript, and a hiring recommendation. Everything needed to decide in five minutes.

Why TA teams switch to SelectPrism:

✅ One platform replaces your screening tool, interview scheduler, assessment software, and proctoring suite

✅ Recruiters stop managing calendars and start closing offers

✅ Every candidate gets a fair, consistent evaluation regardless of when they apply

✅ Hiring managers only meet the people who actually cleared the bar.

We're excited to share this with the PH community and would love your honest feedback. We'll be here all day answering questions.

🎁 Product Hunt Launch offer: 50 free AI interviews!

👉 Try SelectPrism: https://selectprism.ai/agentic-ai-interview

Thanks for checking us out, and huge thanks to our hunter @chrismessina for hunting us 🙏

we just went through a hiring round and the screening alone took longer than the actual interviews. if this actually catches the candidates you'd shortlist yourself without losing the good ones to slow pipelines thats a real solve

Love the focus on solving operational hiring bottlenecks. Curious whether the biggest impact is for enterprise hiring teams or fast-growing startups?

The recruiter pitch is clear, but the part that makes or breaks tools in this category is the candidate side. People can usually tell within 30 seconds when they're being interviewed by an AI, and how that lands quietly shapes whether they accept the offer later.

Do candidates know upfront, and have you seen the disclosure affect completion rates?

Loved the design and UX from the website, but had few more queries.

1. Do i need a separate ATS, or does this already have an ATS inside it? (to track the status of candidates)

2. Will these free 50 interviews have all the features enabled, or will their features be locked behind other tiered pricing?

3. Will there be bulk export options for all the interviews that we take from this platform?

4. Is it possible right now or in future, if i can just upload the resumes in my chat (be it slack or teams) and just ask the bot to match it against specific job that i have already posted on the platform?

5. Will all my jobs be public like naukri.com or will the job applications/interview would be invite only? (like i don't want it to be public as there can be too many applications)

6. How will the scoring work? Like, is there any way i can change the weightage of scoring, different HR/Project managers have different prioritization on what is more important in a candidate?

7. Do you give whitelabling for this?
   
   

One thing I keep noticing in hiring is that great candidates often get filtered out too early because resumes only show a tiny part of someone’s actual potential.

But does SelectPrism focus more on pattern recognition from past hiring success, or is it designed to uncover candidates who might be unconventional on paper but strong in real-world performance?

The 60,000-skill taxonomy built on 60M+ job descriptions is the technically interesting differentiator here. But the hard part isn't the taxonomy size, it's keeping it current. Skills decay and emerge faster than any static dataset can track (a skill that was senior-level two years ago is now table stakes). How does SelectPrism handle taxonomy freshness? Is there a continuous ingestion pipeline from new job postings, or does the graph get periodically retrained on a fixed schedule?

Really cool launch 👏

The biggest hiring pain, honestly, is how much time gets lost in screening and scheduling before teams even talk to the right candidates.

Curious - how does SelectPrism evaluate softer things like communication, confidence, or leadership potential? Especially since those can look very different across roles and experience levels.

Used this on a real pipeline, not a test run. The time between 'JD uploaded' and 'shortlist ready' genuinely shocked me. Solid product, @thecleric Congrats on the launch! 🎉

Using agents for initial screening and interviewing targets the right bottleneck. Most hiring funnels clog at resume review and first calls, not at offer. Does the agent adapt its questioning based on role type, or does every candidate get the same flow regardless of seniority?

The Product solve a big challenge in hiring process the user interaction was good. Just wanted to know how do you tailor your interviews as per different job requirements and bring the feel of a real interviewer interviewing you

Congrats on the launch! The skill-graph parsing and anti-cheating checks look incredibly robust. Quick question on the ATS integrations: do you currently support GetKnit out of the box, or is that on your upcoming roadmap?

Congratuations @thecleric for the launch. Having spent my last 5 yrs in HRtech space, can say that this product feels like well thought and crafted to remove lots of hiring hustles. Best of luck !!

I'm not a car guy and have a lot of respect for anyone reinventing an iconic brand like Ferrari. But this looks a little cheap, doesn't it? It's giving toy car vibes.

Perhaps I need to see it in person. I'm sure it's FUN to drive.

Impressive work from @mike_matas and the LoveFrom team (this is where Jony Ive's Apple Car work must have ended up!).

Kind of wild that the promo video doesn't even show an energetic test drive — it's only focused on aesthetics and that interior, which is kind of surprising given the torque electric cars typically offer!

The "tactile controls uniting physical and digital" line is the most interesting product decision in the whole reveal. The industry's been racing toward touchscreen-everything for a decade, and the driver backlash has been quietly building — even reviewers of other EVs keep calling it out.

Curious: which interactions stayed physical and which went digital? That breakdown alone would be worth its own write-up.

Curious how the EV transition lands at Ferrari pricing tiers — charging-infrastructure economics look very different at supercar volumes. From the renewable-energy deal side (I maintain a set of Eloquens models on project finance), the interesting question isn't whether EVs are coming but how grid and charging buildout for high-end EVs sits alongside the same megawatt that's quietly powering data centers. Beautiful car either way.

The design looks flawless.

It's really interesting how the launch materials are focusing so heavily on the interior and life on board rather than just raw 0-60 acceleration.

the voice input angle is interesting but voice to structured data has a confidence problem. people speak in fragments, change direction mid-sentence, use filler words. curious how the AI decides what's signal versus noise when someone rambles their way through a form field

100+ templates is a lot, I really want to know what has been the most popular use case of this product so far.

Congrats on launch day! The free tier with a generous quota is a good call for a product like this.

the “humans don’t communicate like APIs” angle is strong. i’d be curious how you handle cases where the messy answer contains ambiguity that should not be silently cleaned up, like “maybe Tuesday unless Sam replies” or “use the old address for now”.

do you surface that as a confidence/review step somewhere, or does the form owner define which fields are allowed to be inferred vs. need explicit confirmation?

The “last-statement wins” rule is smart, especially because it matches how people usually listen. The part I’d be careful with is fields where the correction itself is useful context, not just noise.

For example, in sales or hiring intake, “actually no, use Wednesday” may be the final answer, but the earlier Tuesday mention can explain availability, urgency, or uncertainty. I’d love to see a lightweight audit trail for high-impact fields: final structured value, confidence, and the snippet that caused the value to change. That would make messy input feel safer without forcing everyone back into rigid form behavior.

Hello from Greg, Max and the Parsewise team!

Having seen Parsewise provide tremendous value in production for our enterprise customers (incl. UBS, Compre Group, Thinksurance), we are excited to launch our API!

The Problem
Today, you have to build and maintain complex document processing pipelines with changing business rules. You parse, classify, and rely on structured responses from LLMs or IDP tools (e.g., Reducto) to get individual extraction results that you piece together with other bits of information. There’s no reliable way to catch when information contradicts itself, which is risky. Finally, you build a custom verification UI for your operations team to deal with LLM mistakes.

The Solution
We provide an API to abstract all that away into a single call. You provide multiple documents and the desired output to get back a response with resolved values, flagged contradictions, and full traceability across documents / pages that you can display in your own app.

Get Started
=================================================
Sign up with free credits: https://www.parsewise.ai/get-started
(use the Agents.md for a 1m integration)
=================================================

contradiction detection across documents is the part most people build manually and badly. when your agent pulls data from 3 sources and they disagree, knowing which one to trust without a human checking every time is a hard problem. how does the lineage tracing work in practice

How does it handle documents with different formats in the same call say a scanned PDF, a spreadsheet, and a plain text file? Does accuracy drop when mixing formats?

The lineage-down-to-bounding-boxes is the part I'd actually pay for — most extraction APIs hand you a value plus a confidence score and leave you to trust it. But bounding boxes only make sense for visual docs; when the source is a spreadsheet or plain text, what does "lineage" resolve to — a cell reference, a line number, nothing? The human-validation UI only works if every value points back to something clickable.

This caught my eye because tokenization is the silent killer for that language group. Mandarin and Thai have no spaces, Japanese mixes three scripts in one sentence, Korean spacing rules are loose. Did you build the segmenter yourself or wrap something existing like jieba or MeCab? Great job!

This is pretty cool IMO. If we could turn any content we consume and face to in every-day life into a language we wanna to learn, we would be native speakers soon!

This is cool! I studied languages (French and Italian) at uni and find the way that we learn languages fascinating. Learnt Mandarin at school too, which I found much harder to pick up much to Mrs Pan's dismay, so something like this would have really helped! Does it work with any video, do you just select a video and it adds subtitles?

I think situational learning is much more important. For example, I arrive in Mexico and go to a restaurant. I take photos of the menu, dishes, objects on the table, and so on, and it immediately shows me the names using real examples. That would be really cool to build - I would use something like that myself!

This is actually a very cool idea. Most people already spend hours watching Netflix or YouTube, so turning that time into language learning makes a lot of sense. The vocabulary tracking feature looks especially useful for staying consistent without making learning feel like homework.

Hey @chrislabonty I’ve been waiting for a tool like this for a long time. I don’t really enjoy reading subtitles through every scene, but I do enjoy learning new languages. Learning through a different format like this sounds much more engaging and useful for me. Keep up the great work!

Love the concept behind LangPanda. Learning through shows feels much more natural compared to traditional language apps, especially for building listening skills and real vocabulary. The instant dictionary and flashcard creation are a really smart addition. Congrats on the launch!

I'd love to see it in action! Do you have demo content? I couldn't find it on the home page.

Every brainstorm, every breakthrough, every good prompt I had with ChatGPT, Claude, Gemini, or Grok lives inside a vendor UI. A week later I cannot find it.

Kept captures my AI conversations the moment they happen and saves them on my own machine, where I can search them, link them, and build on them.

Features:
- Auto syncs your AI conversations as they happen, straight from the provider
- One .md file per conversation (Obsidian compatible)
- Built-in Agent runs over your own vault, not a cloud index
- Knowledge graph rendered in 3D, surfaces forgotten threads
- BYOK (OpenRouter / Anthropic / OpenAI)
- Additional OpenClaw and Claude Code MCP support to scan your vault

Install is a one-minute job: run the desktop installer, then drag and drop the browser extension into Chrome.

Open source, MIT licensed. No account, no cloud, no subscription. We would love to hear your feedback.

Supported: ChatGPT, Claude, Gemini, Grok, Kimi.

Try it: https://kept.work

Local Markdown is a great default. The next hard thing, I think, is deciding what should graduate out of the chat archive.

A raw AI conversation has a lot of dead ends and temporary context in it. The really valuable pieces are usually decisions, reusable examples, prompts that actually worked, and constraints you don’t want future sessions to forget. I’d love to see Kept make that boundary explicit: archive everything, but help people mark which parts are “memory,” which are just transcript, and which are stale enough to stop influencing future work.

Hi Tibor, congrats! the missing piece! markdown + obsidian-compatible was always going to win. when does the knowledge graph beat full-text search in practice? "where did i decide X" feels like full-text. "what path got me there" feels like graph. is that the split? good luck with your launch.

Local Markdown feels like the right default for AI chats because the valuable part is usually not the chat UI, it is the reusable context you want to move into notes, docs, or a repo.

The detail I’d be most curious about is export fidelity: do you preserve model name, timestamps, attachments, code blocks, and source links in a predictable frontmatter/schema? That would make it much easier for teams to treat chats as a durable knowledge base rather than another archive.

Local-first for AI chats just makes sense. The vendor lock-in is real, you have a great prompt and a week later it's gone inside some UI.

Did the same building a finance app, kept everything on-device instead of a server. The annoying part was syncing across devices without running my own cloud.

How are you handling that with a local vault? Or is it single-machine for now?

Capturing across ChatGPT, Claude, Gemini, Grok, and Kimi means riding five web UIs that each redesign on their own schedule — is capture a browser extension scraping the DOM, or something more durable? That's the maintenance treadmill that kills these tools: a provider ships a UI refresh and the archive silently stops capturing until I notice a week of gaps.

The "understands Django and FastAPI ORM relationships" claim is the interesting one — does that come from reading my models.py statically, or does it need a live DB connection to know the actual schema? Generic AI tools fall over right here: they'll happily generate a query against a relation that doesn't exist. Curious how deep the model-awareness goes before it starts guessing.

I think for a beginner/intermediate user like myself, this would be a great tool. I usually switch between Windows for personal use and Mac for personal/work, so a browser-based, beginner-friendly Python tool would be invaluable. Keep up the great work!

Interesting angle focusing specifically on the Python stack instead of trying to be an “everything IDE”.

Feels like more AI tools are winning now by going deeper into a niche instead of broader.

Framework-specific AST awareness for Django and FastAPI models is what separates this from generic AI tools. Most don't understand ORM relationships or dependency injection patterns. We've wasted hours cleaning up AI suggestions that looked plausible but broke with SQLAlchemy migrations. Does marpy maintain cross-file context, like tracking model schemas and router dependencies across the full project?

who is the primary user here. a Python developer who already has a local setup and VS Code configured is probably not switching to a browser IDE regardless of how good the AI is. the person who might actually love this is someone who knows Python but hasn't gotten deep into tooling yet. are you building for the experienced dev or the intermediate one because those are pretty different products

A Python-native browser IDE makes sense if it reduces the setup friction between idea, backend, and deployed app.

The ephemeral spin-up-run-terminate model is the right abstraction for batch ML jobs. We've burned significant budget on GPU instances idling after failed training runs, especially when a job crashes at epoch 40 and the instance just sits there. How do you handle mid-job failures and artifact persistence? Does the runner automatically sync outputs to S3 before terminating?

it is a cool problem to solve we also have similar kind of problem at our backend we are just paying for idle server times on aws, but one query is like where are you keeping the instance output and is it also available for cpu

Hi everyone!

MiniCPM5-1B is currently the strongest open-source model under 2B for on-device use:

It hits SOTA in the 1B-class on agentic tool use, code generation, and tough reasoning tasks while keeping a very small footprint.

The INT4 weights are only around 0.5GB, which makes the local story much more real.

OpenBMB also shipped a cute Desktop Pet fully powered by this model — completely local, no cloud:
https://www.youtube.com/watch?v=Ee0slMW8SEk

SOTA at 1B parameters running fully on device is wild. the cost of not needing cloud inference adds up fast when you're running agents all day. 131K context on edge hardware is the part I'd want to stress test

Congratulations! The agent learning your tone from feedback and carrying it across messages is the feature that will make this sticky over time. Really well thought through!

Does this operate from a terminal or a web app?

Congratulations on the launch, much needed! Upvoted :)

I have been using Apollo/Prospeo + Clay + ChatGPT as my stack, and it has been hard moving around tools.

So do agents connect to these tools? What accesses would they need for this?

Congratulations! 🥳

For reviewing drafts, do you support some kind of Slack/Telegram/etc bots or is it only on the website? It might be a nice convenience/quality of life feature. Good luck!

Hey everyone! I'm Olivier, one of the creators of blokdots.

With 3.0 we rebuilt everything from the ground up. We replaced Firmata with our own lightweight C++ framework and serial protocol — which finally unblocks us to build the things we always wanted to build.

What that unlocks in 3.0:

• Standalone mode — upload your project and run it without a laptop connected

• C++ code export — hand off to engineering or continue in the Arduino IDE yourself

• Outputs as triggers — let actions fire from outputs, not just sensors

• Multi I2C support — use multiple I2C components simultaneously

And this is just the beginning! Having full control over our stack opens up a lot of exciting directions we're already working toward.

We've been teaching and using blokdots in real design workflows for years, and the gap between "I have an idea" and "I have a working prototype" is still too wide. This is our ongoing attempt to close it.

Core features are free. Would love to hear what you build. 😊

Congrats on the launch! 🟨 Looks amazing!

this is really cool, congratulations on shipping! the figma + protopie hook is the real move! when a designer changes the digital prototype, does the hardware know? one-way (hardware fires, figma reacts) or bidirectional (figma state flows back to motors and LEDs)? congrats on your launch!

Looks like a great product
Congratulations!
Also, how reliable is standalone mode for longer-running prototypes like running a demo loop at a trade show for 8+ hours?

Interesting to see a platform focused specifically on the Python ecosystem instead of trying to support everything at once. The developer workflow focus and AI-assisted productivity angle stood out to me.

Congrats on the launch! Sharing this with my maker friends at the fab lab.

Since this shares a foundation with your Terraform provider — where's the line you'd draw between reaching for the CLI versus just declaring records in Terraform? Anything that lives in version control I'd default to the provider, so I'm trying to picture the CLI's home turf: one-off changes, incident response, scripting renewals?

Secondly, The "structured output for CI" line is the part I'd lean on hardest. Is that stable, documented JSON I can pin a script to, or is it formatted-for-humans output that'll shift between releases and quietly break my pipeline? CLI tools live or die on whether their machine output is a contract.

Overlaying comment spikes against YouTube's own Most Replayed data is the smart move — those two signals don't always agree, and the gaps are probably the interesting part. When a moment is heavily Most-Replayed but the comments are quiet, or vice versa, does LikePulse surface that divergence, or just show both tracks and leave me to eyeball it? The disagreement is where the real insight usually hides.

Really interesting concept. Most analytics tools show views and retention, but LikePulse focuses on audience reaction moments, which feels much more actionable for creators and marketers. The engagement heatmap combined with AI explanations is a very smart touch. Congrats on the Product Hunt launch

the product detection feature is the one that feels slightly out of place with the rest. heatmaps and key moments are clearly for creators and researchers. amazon product detection feels like a different user entirely, affiliate marketers or brand analysts. are those actually the same person in your head or did that feature come from a different use case you're testing. curious because it changes who you're building for pretty significantly

knowing WHERE in the video people react is way more useful than just total view count. this is basically a free focus group for every video you publish

Really like the framing of overlaying comment spikes with Most Replayed — the two signals say different things and the gap between them is where the interesting stuff lives. I run the Mod3Loop YouTube channel on financial modeling, and the moments people comment on are almost never the ones I'd predict from watch retention alone. Tools that surface that mismatch quickly would change how I edit. Following.