PH热榜 | 2026-04-20

Hi! Really excited to finally share what we've been building.
I’m the founder of Project Mirage. We are a team of Designers, Developers and Engineers who have been building in Consumer Hardware for over 8 years, now building in AI Interfaces.

What is Dune and why is it called so?
Dune is a context Aware Keypad for mac that reads which app is in the foreground and automatically changes what its 3 keys do based on what you're doing. We call it Dune because a sand dune is never one thing. It shifts, quietly and constantly, shaped by whatever surrounds it.
That's what these three keys do. They observe what you're doing and become what you need, right then.

What Makes Dune Different
Dune is context-aware, meaning, its keys update automatically based on the app you are running. Unlike other keypads that lock you into setting up keyboard macros for each app, Dune comes with the most-used commands and complex workflows already built in. It is also highly customizable - you can write your own scripts and connect your own agents to trigger via Dune. It reads your active app and surfaces the three most relevant actions automatically, in real time.

The Problem & Our Solution
The way we interact with computers hasn't meaningfully changed in 45 years. We still rely on browsing through screens, clicking multiple links, and memorizing unnecessarily complex keyboard shortcuts for everyday actions. Most shortcuts are buried, forgotten, or simply never discovered, and that's not a user problem, it's a design problem.

Meanwhile, what we actually do on computers has gotten far more complex - developers juggle dozens of tools at once, and meetings now run back-to-back with controls scattered across cluttered interfaces. The friction is constant, and it adds up.

We spent months experimenting with new interface paradigms, showcased three of them at CES, and built Dune to put the best of what we learned directly in your hands. A three-key Mac keypad that reads your active app and surfaces the right actions automatically, whether you're coding, in a meeting, or getting things done.

Features & Benefits
1. Context Awareness: Dune detects which app is in the foreground and automatically updates what its 3 keys do in real time, so you never have to manually switch profiles or reconfigure anything.
2. Instant Actions: Every key is always mapped to something relevant to what you are doing. In GitHub, that means raising a PR, approving/rejecting a change. In your meeting app, joining a call, toggling your mic and controlling your camera with one tap - all while juggling a hundred different tabs.
3. Calendar Sync: Dune syncs with your calendar so you can join your meetings in a single click. Works with Zoom, Teams and Google Meet.
4. Agent Triggers: Trigger your AI agents or agentic workflows directly from your desk. An email assistant, a calendar agent, or anything else you've built in Claude can be activated from the same three keys without switching context.
5. Custom Macros and URLs: Connect any keyboard macro or URL in the Dune app and define exactly what each key does across any app or workflow.

Who Is Dune For?
Dune is built for anyone who lives on their Mac.
1. Developers: Approving a PR on GitHub takes 4-6 clicks on average. Multiply that across a full day of reviews, commits, context switches, and agent triggers and you're spending more time navigating your tools than actually building. Dune maps its three keys to the actions you reach for most in GitHub, VS Code, Claude and more, so you stay in flow.
2. People who live in Back-to-Back Meetings: One tap to join a call, a dedicated mic toggle that auto-brings your meeting window to front, and a camera key so you're never fumbling at the wrong moment. Syncs with your calendar and works with Zoom, Teams and Google Meet.

We'd especially love to see what you build with it.

Drop a comment! I’d love to hear what shortcuts have been frustrating you and share how Dune can fix that for you.
---

🎁 Product Hunt Launch Offer
33% off for our early birds, use code PRODUCTHUNT99 at https://www.projectmirage.ai/order

⭐️ Important Links
Site: https://www.projectmirage.ai/
Setup Guide: https://boom-cuticle-650.notion....
All FAQs: https://www.projectmirage.ai/#du...
Demo: https://www.youtube.com/watch?v=4Hn_ece7NVc

Is it reading the active window process or something deeper than that?

Early user here - thought it would take weeks to build the muscle memory. Took about two days. Now if the keys are not there I notice immediately. Nicely done.

Is there a way to temporarily lock the keys so they stop remapping?

Three keys in a vertical stack rather than horizontal is interesting. Easier to reach without moving your hand off the home row.

Curious what the latency looks like between the macOS accessibility API firing and the key remap completing.

What is the loudest the keys get? Hot desk life is real, and some of us have to be considerate.

This is cool. Context-aware is doing a lot of work in the pitch. Macro keypads usually trade a few saved seconds for remembered mappings so net productivity ends up a wash. How does Dune's context trigger? Active window, calendar state, time of day.... ? Congrats, and good luck :)

Congratulations. And happy product launch. @apoorv_shankar

Cool project! Guys, you're great!

Congratulations!!!

Interesting contrast to tools like Stream Deck fewer buttons, but smarter ones. Feels like a bet on intelligence over customization.

Curious how fast and accurate the app detection is in practice. Even slight lag could break the flow you're aiming for.

Okay, I love this! I'm really excited. The vertical stack of keys looks like a smart idea. How is the latency?

Hardware that actually thinks about the person using it. That is what we were going for from day one and I think we got there. Really proud of this one.

custom scripts alone are worth it. i have sooooo many workflows I want to build with this. congrats on the launch, really proud of what we shipped :)

Waiting for an update wherein it connects via bluetooth and I could use it as my external Knuckles. Kudos to the team!

How long does it take to learn a new app? Like if I add a tool to my workflow tomorrow, does Dune pick it up automatically or do I need to do something?

the fact that it changes based on the foreground app is sneaky smart 👀 been losing my mind switching between cursor claude and meetings lately so this solves the exact right problem.
how hard was app detection to get reliable across all the chaos?

This looks fantastic! Congrats on the launch! It's USB-C connection, but you have M2 or later on the site for compatibility. Any specific reason an M1 Mac won't work? There are 2 USB-C ports in the same location as the newer models

How fast is the remapping actually? Like is it instant or is there a moment where the keys are in between states? Never clocked it properly.

Designing something this compact while keeping it intentional and complete was a unique challenge. Really Happy with how it turned out!

Every surface on this was intentional. Good to finally have it out in the world!

As someone who has spent years building hardware, shipping something this precise and this smart at the same time is genuinely satisfying. Really proud of what this team pulled off.

Three keys sounds like nothing until you realize those three keys are always the right three keys. This is a good one! 😎

nice, i basically wanted to do the same with the elgato custom stream deck module and call it "clawboard.engineer", but this is sooo much better. i need this in my life.

Why i need this ? If I can also use siri much faster than this

Looks dope man! Good luck

Congratulations on the launch! Looks exciting, and I love the naming, waiting for more geographies to come :)

three keys seems minimal but probably smart. love that it's built specifically for the GitHub/VS Code/Claude workflow since that's our daily stack. wondering if there's any way to customize actions beyond what the AI suggests, or if it learns your patterns over time?

Hi everyone!

This is a wayyy interesting bridge for Claude into physical form.

What @felixrieseberg shipped here is small, but the idea behind it is bigger: Claude Cowork and @Claude Code now expose a lightweight Bluetooth path that lets maker hardware react to sessions, recent messages, and approval prompts. The reference device is an ESP32 desk pet built on M5StickCPlus, which honestly feels like a very fitting choice — M5Stack really is one of the friendliest entry points in the hardware world🔥

It already makes interacting with Claude Code more fun to look at and touch:

But the interesting part is not just approvals. What happens when this lives on an e-ink kanban? Or when the hardware surface is something much more ambient, tactile, or playful? This is clearly the start of a much bigger hardware interaction layer for Claude!

@zaczuo BLE from Claude desktop to ESP32 is clever — are permission prompts forwarded over BLE as bidirectional events, or is the hardware side purely read-only status display? The approve/deny button use case would be a great hardware interaction for Claude Code agents.

Does it work on my GameBoy Color. thats all I want to know :D

oh this is wild — BLE straight from the desktop app is such a fun primitive. any plans to open up the event stream beyond permission prompts? would love to drive a little status light off tool calls

will you also bring back /buddy? @felixrieseberg

Most team track usage , but not what actually makes it to production. This kind of visibility could really help cut wasted spend . Curious if it also highlights why some AI generated PRs don't get shipped?

Hey Product Hunt 👋

I am Alex, founder of @Waydev Nine years of building engineering intelligence. I have never seen a shift like this one.

AI agents are writing your code. Nobody audits the output.

4% of public GitHub commits are already authored by Claude Code. Companies are spending up to $195 per developer per month on AI coding tools. Almost none of them can prove the spend is working.

That is the gap we rebuilt Waydev to close. The new platform measures the full AI SDLC:

• AI Adoption — which tools your teams use, what you spend per vendor, per team, per repo

• AI Impact — follow AI code from IDE to production. See where it ships and where it dies

• AI ROI — cost per PR, cost per shipped line, tokens consumed vs code shipped

• AI Checkpoints — commit-level attribution. Which agent, how many tokens, what percentage was AI

• Waydev Agent — ask anything. Closes the loop by feeding insights back to your AI through MCP

AI adoption was the easy part. Proving what AI actually changed in production is the hard part. That is what we built.

In the comments all day. Ask me anything.

— Alex

This feels super relevant right now. A lot of teams are thinking about this problem. Will give it a shot.

this is exactly what we've been missing. we use Cursor and Claude Code daily but have zero visibility into which suggestions actually make it to prod. the cost per shipped PR metric is brilliant - finally a way to measure actual AI ROI instead of just "feels faster." curious how the agent tracking works across different IDEs?

This hits a real blind spot. Everyone is adopting AI coding tools, but almost no one can tie usage to actual shipped value.

nice way of looking at your team's output, now together with visibility for generated code. will try it out soon.

love that you're tracking acceptance rates by vendor. we've been debating Copilot vs Cursor internally and it's all gut feeling right now. being able to see "Cursor had 73% acceptance but Copilot code shipped 2x faster" would end those arguments quickly. does it handle when devs modify AI suggestions before committing?

Congratulations for this release, I know how much work you and the team put into it. Now, this version looks like a very robust solution, love it! Can't wait to plug the new Agents into our workflows and see what we actually ship 🫡

Finally something that looks at actually measuring productivity beyond just lines of code. With AI agents, generating code is becoming the easy part, but the more important question is what actually makes it through review, ships to production, and creates durable value. Otherwise we risk confusing velocity of spitting code with actual progress.

This feels like the right lens for understanding AI’s real contribution to engineering teams. The one question I'm still trying to figure out and I'd love your perspective: how do you connect these engineering metrics (output) with the business KPIs (actual business outcome)?

measuring actual shipped % of agent code is such a missing metric — everyone's tracking tokens spent, not outcomes. how do you handle attribution when a PR goes through multiple agents + a human review before merging?

I’d love to know if your platform highlights trends or just raw metrics

Is this for big enterprise or even for small startups?
Also I didnt find the pricing model. Not sure what I missed.

Looks really cool.

How do you compare against https://macroscope.com/ ? I like 1) their github integration and the code suggestions, 2) the the sprint analysis.

Token consumption tracking is interesting — how does that work in practice with something like Claude Code, which runs autonomously and can spin up multiple sub-agents mid-session? Are you capturing tokens at the session level, per file touched, or per PR? The attribution question gets messy fast when one 'task' spawns 40 tool calls across 3 agents.

Hey Product Hunt — Josh here, founder of kaizen.

This launch means a lot to me because I’ve been building versions of this idea for almost 10 years.

It started with my own experience as a runner. I ran my first marathon at 17 and followed the kinds of plans most runners follow, but I kept running into the same problem: they only really worked if life went perfectly. Miss a few sessions, get a week wrong, and the whole thing could start to feel off track.

Over time, I became more interested in what actually helps someone improve in the real world. Turns out consistency is the primary determinant, closely followed by the ability to gradually increase training load. So I developed a system that ingested all my recent running history (relative to my past races and training) to understand how fit I was today, and work out how much I'd need to run each week to achieve my goal.

That system changed my own running completely. I went from a 3:24 marathon to 2:28 and became the top-ranked U23 marathoner in the UK. But more importantly, running stopped feeling like a cycle of falling behind and starting over. It became sustainable and enjoyable.

That’s where kaizen came from.

Really excited to finally share it here, and grateful to everyone who’s helped get it to this point.

Been using kaizen for a few months now and generally really enjoying it! Sometimes I find the mileage goals are a bit tooo high (but I'm not generally a high mileage runner), but honestly, the predictions are pretty accurate (I ran a tempo half marathon a few weeks back and it had the pace spot on). I really like that you can plan your week to hit your goal, and the adjustments you get after each run to your goal time. I also like that you have a predicted time for all the distances (they're not just focused on the marathon, like half of the running community seems to be at the minute!) Also a big fan of the "Tailor your prediction" function - I used it just this morning and now I feel my marathon predicted time is (hopefully) a lot closer to reality!

I work with a coach, so kaizen is more of a complement to my training / let's me see if I'm on track for my goal, but I am a fan and would recommend (I just wish I knew a little more about the science behind it and how it calculates its figures!)

Lining up to the start line of

Boston Marathon today with a good grasp of my finishing time thanks to Kaizen! Love the flexibility!

congrats on the launch! the adaptive angle stands out. Most running plans are fixed 16-week blocks that punish you for a bad week 7. How does kaizen decide when to push versus back off? Pace trends, RPE, missed sessions, or a mix?

As someone who has always preferred to run freely based on how I’m feeling or sometimes constricted by busy days and weeks, I struggled to follow prescribed training plans, as sometimes they would end up feeling like an extension of work. One more to-do in the calendar. This app finally solved that issue, and it quietly become my main training driver. Couldn’t be thankful enough for it.

No more skipping running with friends because I have a specific workout to do.

No more big question marks about what sort of shape I’m in and having to test it or prove it in workouts.

The more I used the app the more I realised I was back to my initial years of running freely and enjoying it.

And when I go to a race I know what my fitness is translating to in race times (it surprisingly always gets the prediction extremely close!).

I even tested not doing anything close to potential “race pace” for months on end and then raced a 10k simply on feel and voila: it matched the pace that Kaizen was predicting for that distance (that I hadn’t run before in training).

Having used Kaizen (and a host of other running apps) it's definitely one I recommend to others.

The training plan gives you the flexibility to run when and how you like. For me, this allowed me to listen to my body a lot better, pushing on days I felt fresher, and taking it easier when I needed the rest. This isn't something I've ever seen in other apps (and I've no doubt is part of the reason so many other apps are driving people into injuries)

(As I understand it) The weekly targets are built on a solid scientific and data led basis, which means I can be confident that the goals I've set are going to be hit. Literally just last weekend I met another Kaizen subscriber who ran a new PB at ParkRun having used Kaizen in training.

This hits a real pain most running plans assume perfect consistency which which just isn’t how life works.

Love it. Use it. Even paid for it.

Hey Josh! It's super cool. I'm currently using another app and I can really feel all you're expressing here. My own plans broke when life happened and needed something like kaizen. Wish you all the best here!

If running is your primary sport, Kaizen is the way to go!

Balancing all the training for long distance running can become a constant puzzle with everything else going on in your life. I think Kaizen really helps simplifying training by making you focus on what matters - gradually increasing your total training load step-by-step.

The dynamic tracking of this training load really helped me see when I was at risk of overtraining. Am convinced this has spared me some overtraining injuries in the past year.

As somebody who struggled to follow my Garmin training plan fr this reason, this is an awesome idea. I'm highly partial to at home fitness apps that adapt with you instead of being generic but claiming personalized. Does it adapt from the first run you do with Kaizen or is there a "learning" period to learn a users habits?

Love this. How does it compare to Runna?

the adaptive piece is really interesting - most training apps just expect you to stick to the schedule no matter what. how does it actually adjust? like if you miss a few days or run longer than planned, what kind of changes does it make to keep you on track?

Kaizen is a hard app to recommend, its ui is slow and clunky, the plans that it gives you are confusing and the price is pretty steep for something that has these issues.

The app tells you how far you should run in the week and lowers the amount of distance if you've had a particularly hard run (tempo), there are ways to say you are planning to run x time at y event but there is no specific plan like runna or garmin it just tells you to run distances per week.

If your looking for future weeks they dont update unless you plan runs in their excruciating interface - and even then the distance doesn't seem to be affected until the week is complete.

Id say this app is good if you just want a weekly goal and no specific session runs i.e. 3x5 at 5km pace, if you want to do that you have to do it off your own back and setup your session yourself.

As i say difficult app to recommend to anyone really.

P.s. i did try and email them some suggestions and they told me that they were working on some of them, now that they've officially launched on product hub i can see that it was not true.

I wouldn't re-sub to this app.

Hi, Jae here. I’m the CEO of @TwelveLabs

Today, we’re launching Pegasus 1.5, the first video language model turning video into queryable data assets. What would you build if your video was as queryable as text? Try the free Playground: twelvelabs.io

Video is the most opaque data source; it’s hard to know the content of your video without simply watching the video. Pegasus 1.5 lets you understand your video library, autonomously, on-the-fly, and at scale. More than that, it future proofs your archive and enables agents to actually navigate it with enriched and custom-defined metadata.

What’s New:

1. Time Based Metadata: Generate custom, time-coded metadata, based on your exact need. Some examples include: segment every time the speaker changes, segment every time my favorite basketball player dunks, and segment every time my logo appears on screen.
   

2. On the Fly Processing: Start with just one video, and get value immediately. If you’re a creator who needs to chapterize your content for youtube, with transcription and key events, upload the video on TwelveLabs, and Pegasus 1.5 will give you exactly what you need.
   

3. Multimodal Prompting: Pass in an image, and tell the model to show you every time the object in the image appears. Try it for product placement or for tracking your favorite player across a game.

We're proud to make a model that actually helps you understand your video content, in the way you want. We outperform top general models on segmentation and on multimodal inputs. We support 2 hours of video, which is more than two times other models. And, we’re way more cost efficient. Check it out, and would love your feedback!

I had a blast collaborating with @emilykurze and the @TwelveLabs team on this launch.

Pegasus 1.5 is a significant leap in generative video AI: autonomous and reliable segmentation, long-form video support (up to 2 hours), with SOTA performances (30% better than Gemini 3 Pro, 3.1 Pro, and 3 Flash).

Try the free Playground at twelvelabs.io - Looking forward to see what you're building!

Impressive to see long-form support up to 2 hours. That’s where most current tools struggle, so this feels like a meaningful improvement.

Hey Product Hunt - Giorgi here, founder of Tasteit.

This one is personal.

I grew up in Tbilisi, Georgia - a city where every important moment in life happens around a table. Business deals, friendships, love stories - all of it starts with food. I never understood why there was no platform built around that.

Every social network we use today was built around something else - photos, status updates, professional profiles. Food, the most frequent social behavior in human life, never got its own home.

So we built it.

Tasteit is the Food Social Network. A dish is your entry point. You discover it, see who loves it, and turn that into a real meal together. No awkward intros. No profiles to fill out. Just food as the most natural reason to meet someone.

We started in Tbilisi. Then Berlin. Then New York. People in 70+ countries found us without us looking for them.

Today we're launching here as part of The Pitch by Deel global competition - and it means a lot to us.

Really grateful to everyone who's supported this from the beginning. And excited to finally share it with this community.

hands down one of the most amazing consumer app release I have seen in very long time.

Love this !!

The concept is solid and Tbilisi as origin story is genuinely charming. The real test is whether "a dish you both love" is enough of a wedge against just... texting a friend and picking a restaurant on Google Maps. 👀

Will it help me to find someone who likes pizza with pineapple? 😁

I would recommend people trying this out, I just went to eat out with a bunch of interesting people.

There is noting more interesting than sharing a meal with a stranger and chatting about life. The people who used to travel and stay in hostels know what im talking about. :D

every time I've met someone interesting recently it started with "let's grab food" not "let's connect on LinkedIn." the awkward intro problem is real.

Interesting concept. How are you turning that initial interest in a dish into people actually meeting up in real life?

Favorite app. Literally had a dream where I fought with my family and had no one to dine with but I wasn’t worried because Tasteit exists. I know my foodmate is always a click away

Is this app target to specific region?
Getting the grants is different region - have a very diverse flow. How u are able to manage this

As someone building B2B tools in India, what's the biggest win you've seen so far, and does it handle region-specific programs like Startup India or EU funds?

This solves a real pain point. We looked into government grants for our B2B tech company last year and gave up after the third application — the compliance requirements and formatting were brutal even for straightforward programs. Curious about two things: how do you handle the highly specific technical narratives that grant reviewers look for (vs. generic AI-generated language), and what's the success rate looking like for applications submitted through Granter vs. traditional consultants?

"Fully integrates into your company" — what does that actually mean technically? Does Granter connect to existing tools like email, accounting software, or a document drive to pull company data automatically, or does someone still need to manually feed it financials, headcount, past projects, etc. before it can write a decent application?

Grant writing and government proposal writing have a lot of overlap — both are competitive structured applications where you're fighting for limited awards. One thing I've always found tricky in this space: the AI handles the narrative draft well, but compliance checklists (is every requirement explicitly addressed?) are where things fall apart.

Does Granter do any compliance mapping — like cross-referencing your narrative against the grant's stated requirements? That's the piece I'd want before submitting anything.

Hey Product Hunt 👋 I'm Liina, I've been selling and growing things for decades (yikes, sounds old) some things kept repeating itself. So we built Knowzilla to fix it: people can sit through endless onboarding and still freeze when a real sales conversation gets hard. I ran a top-performing regional team and kept thinking: selling is not rocket science. You need the right answers and confidence, the rest can be automated.

We are a team of sales and engineering leaders - built some unicorns before, now we want to help anyone sell their beautiful idea better.

Standard sales training often collapses under the pressure of a live call. I have spent a large part of my career designing consultative processes and managing sales teams; the primary hurdle is always ensuring that a strategic framework actually translates to a real-time conversation.

I admire that Knowzilla is using AI to solve these problems. Provides the situational awareness required to handle complex objections without the mental tax of searching through playbooks is gold. This looks like a sharp, pragmatic solution for organizations that prioritize precision over guesswork.

Interesting take on the live call guidance problem. We ran a 15-person sales team at our IT services company and the biggest gap was always tribal knowledge — senior reps knew the answers to tough objections but junior reps would freeze. How are you handling the knowledge ingestion side? Specifically curious whether Knowzilla pulls from closed-won deal recordings or if the team has to manually build the knowledge base upfront. That bootstrap phase is usually where these tools stall.

The Playbook creation flow is already a strong highlight for me. The questions you guide users through are thoughtful and genuinely useful, they help clarify key aspects of the product and provide a solid high-level overview. That alone adds a lot of value.

The call assistant is another standout feature. Getting immediate analysis after a meeting is incredibly practical and can save a lot of time when reviewing conversations or extracting insights.

I did run into an issue with the Playground, specifically the email practice feature. When I clicked the button to start, nothing happened.

Regarding the call assistant, I tested it in a language other than English, and the results weren’t as accurate. Improving multilingual performance could be a great area for future development and would definitely broaden the product’s appeal.

Overall, this is a solid and interesting project with a lot of potential. Wishing the team all the best with the launch and what’s ahead! 🚀

Hey Liina! Love the idea! How do you ensure that the library has all the up-to-date information? does it self-refresh or it needs someone (e.g. RevOps) to take care of that?

This is clever. My only worry is that I'd be so focused on reading the suggestion that I'd stop actually listening. Is that just a me problem, or did you run into that during testing? Did it get easier after a few calls? Congrats on the launch!

Curious how this handles long-cycle deals — government BD in particular runs 6-18 months from first contact to award. The "guiding every deal" concept is interesting but most sales AI I've seen is optimized for 30-90 day cycles.

Does Knowzilla have any concept of deal stages that extend that far out, or is it more focused on standard commercial pipelines?

Looks pretty interesting! Is this solution positioned to mainly help SDRs / BDRs and AEs? Or have more technical sellers like Enterprise AEs or Sales Engineers been using this as well? I can image with younger and less technical salespeople this is really useful but also curios to learn how more technical salespeople use it or get value out of it. Keep up the good work!

🌎🫶If you've ever spent a Friday night reprojecting EPSG:4326 into EPSG:5514 and questioning your life choices: this one's for you. Marek here, CTO and recovering geospatial engineer. We built PangeAI so AI agents can finally handle the stuff that made me want to quit grad school. Would love your feedback, especially the brutal kind.

Most AI agents know what they’re talking about, but they have no idea where they are. We’re here to make sure the robots don’t try to take a shortcut through a lake just because it's a straight line on a CSV

Please get in touch with us for info, we'd love to map things for you :)

Johanna here, CEO of PangeAI.

Nowadays everyone seems to be building AI agents…
we’re focused on the ones that actually matter - the ones that understand the real, physical world (and save you from GIS-induced suffering)

Would love your thoughts (I'm German, so honest feedback is basically a love language)

Google just made TPUs a first-class target for PyTorch, and you barely need to change your code.

The problem: TPUs power Gemini, Veo, and the largest AI clusters on earth, but using them from PyTorch required workarounds, framework rewrites, and deep hardware expertise most teams don't have.

The solution: TorchTPU is a PyTorch-native backend that lets you change one line of initialization and run your existing training loop on TPU, no core logic changes required.

What stands out:
⚡ Fused Eager mode: Auto-fuses ops on the fly for 50-100%+ speed gains, zero setup required by user

🐛 Debug Eager: Catches shape mismatches, NaNs, and OOM errors one op at a time so you fix bugs faster
🔁 Strict Eager: Async single-op dispatch mirrors the default PyTorch experience for a flat learning curve
🔧 torch.compile via XLA: Peak performance with full-graph compilation, battle-tested for TPU topologies
📦 Custom kernels via Pallas & JAX: Write low-level hardware instructions without breaking performance
🌐 DDP, FSDPv2, & DTensor supported: Scale distributed training without rewrites
🔀 MPMD support: Divergent code across ranks works without breaking your stack
💾 Shared Compilation Cache: Reduces recompilation overhead across single & multi-host deployments

On the roadmap for 2026:
- Public GitHub repo with docs and reproducible tutorials
- Dynamic shapes support via torch.compile
- vLLM and TorchTitan integrations
- Linear scaling validated up to full Pod-size TPU infrastructure
- Native multi-queue support for async codebases

Different because this isn't a wrapper or a fork; TorchTPU integrates at the PyTorch PrivateUse1 level so you get ordinary PyTorch Tensors on TPU hardware, no subclasses, no rewrites, no friction.

Perfect for ML engineers and research teams running PyTorch workloads who want to leverage Google TPU infrastructure without abandoning their existing codebase.

P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends

honestly the Fused Eager mode is what caught my attention here — getting 50-100% speedups without touching your training loop is pretty wild. been running some PyTorch fine-tuning jobs on A100s and the compile step is always where things get messy. curious how the debug eager mode handles mixed precision edge cases though, that's usually where I spend half my debugging time. the fact that this works at PrivateUse1 level instead of being a wrapper is a huge deal for anyone maintaining custom training pipelines

The Skill Map output is what I want to understand better. Is it a snapshot — like a score you get once after a session — or does it update over time as you do more scenarios? Because a one-time assessment is pretty different from something that tracks how you're actually improving.

Running existing PyTorch workloads on TPUs with minimal code changes is compelling — what's the experience like for jobs that depend on custom CUDA kernels? That's typically where XLA/TPU migration breaks down for large training pipelines.

a pixel art that actually looks good and isn't just a blurred mess... does the ai understand tile-sets or is it just individual assets for now? either way, not needing to spend weeks on sprites is a huge speedup. supported

What inspired DogBase: I'm a SAR K9 handler. My Belgian Malinois Echo started as my service dog during a period of serious illness — and eventually became a fully certified search and rescue dog. During that journey I realized working dog teams are managing incredibly complex operations with notebooks, WhatsApp groups, and spreadsheets. Training data gets lost. Health patterns go unnoticed. Teams have no shared visibility into what's actually happening with their dogs. So we built DogBase — a platform where every session, health event, and performance insight lives in one place, with AI that tells you what the data means. We're not just a tech company. We're in the space. Echo is still active, and his data is in DogBase right now. Would love feedback from anyone working in K9, dog sport, or animal performance — and happy to answer questions about what we've learned building AI tools for operational teams.

"Performance tracking" is interesting for working dogs — what does that actually surface over time? Like, are you looking at things like drive degradation, injury recovery curves, or alert accuracy rates? Would be curious to see what a 6-month performance report actually looks like for a patrol dog.

Hey PH! I soft launched GalaxyBrain here about a year ago. The response was encouraging but it became clear pretty quickly that the foundation wasn't strong enough for what we were trying to build.

So I convinced my smartest friend (@viralplatipuss) to join as technical cofounder and we rebuilt the entire product from the ground up. GalaxyBrain is closer to a spreadsheet engine than a note app, and that level of complexity needed a foundation built properly.

It's ready now. The core is solid. You can have thousands of pages with dynamic values, all connected to each other, all staying perfectly in sync and updating in real time.

It's been a multi-year labor of love. You might find some rough edges on the fringes but the engine underneath is rock solid.

Would love to hear what you think and if this approach has potential.

Hey Product Hunt 👋 I'm Matteo, co-founder of Claro.

We built Claro because we kept losing trust in our own AI outputs. @tameesh and I spent years at Delivery Hero and Yelp trying to wrangle product and supplier data at scale. Tameesh then did NLP/LLM research and published at MIT Review. Every time we plugged an LLM into a data workflow, we got beautiful-looking answers with no way to know which ones were right.

Claro is the AI execution layer for product and supplier data. It takes messy supplier feeds, spreadsheets, and documents, turns them into trusted catalog entities with stable IDs, and keeps them validated and correct over time.

Today we're opening the first public module: Research Agents. 14 task-specific agents that run inside a native table:

📋 Find your perfect list — describe ideal prospects, suppliers, or partners, get an enriched dataset 📍 Build a location list — draw an area on a map, generate POIs with attributes 📄 Capture tables from PDFs — pull tables from reports, pick which to keep 📑 Turn documents into structured data — extract fields across hundreds of docs at once 🌐 Scrape data from URLs — turn a list of links into structured rows 📊 Analyse CSV — upload a spreadsheet, validate and enrich it 🔀 Merge & Map — dedupe and reconcile records across two files and more.

What makes Claro different is what happens to every answer before it hits your table:

→ Multi-model consensus — we push extractions to multiple models in parallel and only accept where they agree
→ LLM-as-judge filtering — a lightweight quality gate catches nonsensical outputs before they reach you
→ Source ranking — first-party and authoritative sources above random blog posts, datasheets above application guides
→ Confidence scores — every cell scored at row, column, and entity level based on source reliability, cross-source consistency, model certainty, and retrieval quality
→ Full citations — click any cell to see the exact passage, URL, or document section it came from
→ BM25 + AI hybrid matching — for entity resolution against your existing catalog, with human approval loops for borderline cases

It's not a chat wrapper. It's validation-first infrastructure with a spreadsheet you can use today.

Three things I'd love today:

• Try it — 200 free credits, no card.

• Break it — throw your ugliest CSV or weirdest PDF, tell me what it got wrong (and what the confidence score said)

• Tell me which agent you'd use first — we're deciding what to deepen next

I'm pitching live at The Pitch Berlin today, so I'll be bouncing between stage and comments, but I'll reply to every one before EOD.

If you run a marketplace, distributor, or multi-supplier catalog and the underlying platform sounds relevant, DM me — we'd love to talk.

Thanks for taking a look 🙏 — Matteo

Hello, Tameesh here, co-founder & CTO 👋

Happy to go deep on anything technical — how we compute confidence scores, how source ranking weights first-party vs open-web sources, how we orchestrate tool calls across extract/classify/generate/search, multi-model consensus, entity resolution with BM25 + embeddings, graph-driven validation, what happens when the model is uncertain.

Short version of what I’m proud of: the confidence score isn’t a vibes check. Source reliability 40%, cross-source consistency 30%, model certainty from output variance 20%, retrieval quality 10%. We run multi-model consensus for high-stakes enrichments and LLM-as-judge as a lightweight filter on top. You can sort a 100k-row dataset by confidence and review only the reds. That’s the difference between “AI output” and “AI output you can use.”

Drop your hardest technical question below 👇

thanks @rajiv_ayyangar for the support as Hunter!

The confidence score + citations on every cell is the interesting part. How does that work across the PDF extraction agent specifically? PDFs can be messy — scanned docs, weird formatting, tables within tables. Does the confidence score drop noticeably on low-quality source material, or does it stay falsely high?

Great for small team/solo shippers. But if someone does not have any written tests, how do we set it up? Do we just explain our product....what is the process?

Idea sounds good. Any plans for MCP so you can teach AI to build the tests when implementing a feature?

Really like the approach with the npm package and pay as you go pricing.

I wonder though why the crow should be better than simply telling Claude Code or Cursor to write some test cases for my project?

@ryanmerket which platform charge $8k for qa per month? What is different in this platform compared to already in the market?

Superb idea. what happens when the test description is under specified ?

@ryanmerket 'write QA tests in plain English' — what happens when the test description is ambiguous? Like 'make sure the form works' — does QA Crow infer what 'works' means from the DOM/context, or does it prompt the user to clarify before running?

"It runs ad-free, requires no account, and keeps all data local, giving users full control." – This is the strongest competitor of YouTube!

Hey Product Hunt 👋

I’m Aditya, the developer behind EchoTube.

EchoTube is an open-source Android YouTube client built to feel fast, clean, and fully private. I started working on it because most video apps today are bloated, ad-heavy, and rely heavily on opaque server-side algorithms. I wanted something that gives a great experience without tracking users or locking them into an account.

The core idea is simple: everything should run on your device. EchoTube uses a fully local recommendation system that learns from your watch patterns - like watch time, skips, likes, and searches - without sending any data to a server.
Key highlights:
• Fast, lightweight UI built with Kotlin + Jetpack Compose
• Clean playback with features like PiP, background play, and casting
• Smart on-device recommendations (no telemetry, no account)
• Ad-free experience with tools like SponsorBlock and DeArrow
• Full control over your data - everything stays local

This is v1.0, so there’s still a lot to improve. I’m actively building and would genuinely value your feedback - features, bugs, or even criticism.

If you care about privacy or just want a simpler YouTube experience, I’d love for you to try it and tell me what you think

How is it different than revanced or morphe .software? Also, wondering how you can keep it legal with Youtube?
Great that there are more options to use Youtube without ads in general!

Open source YouTube client with privacy focus — does EchoTube route requests through a proxy to avoid YouTube tracking the client IP, or is the privacy mainly around avoiding in-app analytics and ads? Wondering how it handles the YouTube login/auth flow.

It's great that there are no ads. Are you considering releasing an Apple OS version in the future?

At Iqana, we’re building the infrastructure that allows investors to access institutional-grade Digital Assets strategies without needing to build a quant team or trading stack internally.

We focus on:

• Systematic, data-driven strategies (momentum, market-neutral, yield)

• Non-custodial setups — clients keep control of their assets

• Advanced risk management and execution infrastructure

In short: we power HNWIs, hedge funds, family offices, and asset managers with hedge fund-level Digital Assets investing — in a simple and secure way.

Happy to answer any questions 🙌

E-commerce accounting automation is a real pain — reconciling Shopify payouts, COGS, returns, and sales tax across multiple states is genuinely tedious. Does CONA handle multi-currency stores and international VAT, or is it focused on US-based sellers first?

Payment Matching is a nightmare indeed... who is this built for — freelancers, SMBs, or accountants/Steuerberater?