I didn't win, but almost every top project had one thing in common: a built-in verification mechanism.

1. Introduction

This is part 2 of my Meta Global AI Developer Hackathon results report.

In part 1, I wrote about why I joined and what I submitted.

In this part 2, I'll introduce the winners announced at the closing ceremony.

2. What I Learned at the Closing

There were roughly 33,000-40,000 applicants, with 20,000 participants.

There were 5,367 submitted projects, with over $1M in total prizes.

At the closing, finalists pitched live and judges reviewed them on the spot.

After deliberation, winners were announced for the Indie and Startup tracks, plus special prizes.

Only the $10,000 Community Vote was still open.

The winner is scheduled to be announced on 2026-11-16.

3. Indie Track Winners

The Indie track is for individuals and small teams.

All five top projects had strong ties to the real world.

🥇 ControlAnything is an AR assistance app for Quest 3. Pinch a real physical device and it automatically generates a working connector for it on the spot.

🥈 Gatekeeper is an inspection tool that monitors new npm packages. It detected malware about a day earlier than public lists, and requires cited evidence for its findings.

🥉 David Barry is a home assessment app for making use of basements. Just take photos with your phone to learn whether you can legally turn it into a second home, and how.

4th: Aurora (Nemesis) is a physics simulation generator for robot training. It builds environments from natural language or images and can retrain legged robots in them.

5th: Video2URDF is a conversion tool that creates robot models from video. It automatically generates URDFs from videos of articulated objects, cutting development effort.

4. Startup Track Winners

The Startup track is for teams aiming to build businesses.

The winners tackled concrete work in the field, manufacturing, and other domains.

🥇 Consai (Punchwalk) is a defect management tool for construction sites. It creates verified punch lists from voice and photos to reduce rework.

🥈 Formwright is a hardware design tool that goes from prompt to physical device. It generates circuits, boards, enclosures, and firmware, all the way to a manufacturing order.

🥉 ORCA is a design support tool for interference between 3D models. It doesn't just detect collisions — it proposes fixes with geometric proof.

4th: Voluma Bio is a 3D measurement tool for research. It automatically generates 3D measurements from micro-CT images to help with analysis.

5th: Kroonen AI (Glimmer Scope) is a tool for improving model interpretability. It visualizes LLMs' internal concepts and lets you probe behavior with dials.

5. Special Prizes

Special prizes went to projects with outstanding technology or creativity.

Several of them also had verification built in.

Most Viral: Adrian's Playtest Agent is a game testing assistant. It plays games from screenshots alone and reports bugs with replay verification.

Best Build on Meta Open Weights: Tessa is an astronomical data analysis tool. It autonomously searched NASA TESS data for exoplanet candidates and found 8 new ones.

Best Agentic: oqed is a code optimization tool. It has an LLM write fast assembly code and only accepts output that passes strict verification.

Best Creative: Team Synthexis's synthexis is an instrument you play with a webcam. You can perform a synthesizer with hand and face movements for intuitive expression.

6. My Result and Takeaways

I submitted Trifecta Checker solo to the Indie track.

It was a checker app that verifies AI output on three dimensions.

The result: no prize, and I didn't make the finalists.

Missing out was disappointing, but seeing the top projects, I understood why.

What many of them shared was verification built into the product itself.

Cited evidence, verification gates, geometric proofs, replay verification, and more were all baked in.

It felt like the question wasn't just what you can make AI generate, but how you make the result trustworthy.

The other thing was how deeply they stepped into the real world.

Construction sites, basements, robots, research images — the targets were specific.

I think my project was close in its focus on verification.

But the link between the target domain and that verification felt weak.

7. Caveats

These results are based on the announcements at the closing.

The Community Vote is not final yet and is subject to change.

The official gallery entries are the source of truth for each project.

The intros in this article are 2-3 sentence summaries of the key points.

8. Wrap-up

Looking at the top projects, I felt that how you build trust mattered as much as flashiness.

Built-in verification and a dive into the field will be my guide going forward.

For my next build, I want to narrow the target to one thing and build it all the way through verification.

References

Closing recording (Vimeo)
https://vimeo.com/1245287583

Official gallery pages for the winners (Cerebral Valley)