Honorable Mentions: 9 Argon Project Favorites
Couple of the winning projects with others chosen for great UI & purpose

Honorable Mentions: 9 Argon Project Favorites

This is Part 2*. Part 1, “One day, 136 projects,” covers the room. Part 2 is about the gallery, and the nine projects I keep going back to.

Congratulations to the winners: 🥇 First Place: polar bears 🥈 Second Place: Morf 🥉 Third Place: eagle. 🎖️ Best Use of Managed Agents: Revive 🎖️ Honourable Mention: Aniket’s Team

They get the attention, and they earned it. This piece is about nine others.

I read the whole gallery, then had Gemini watch the videos. I picked nine. They are my taste, not a ranking, and I did not build a rubric. Still, one habit ran through all nine. Each one let the model be creative and then handed the result to something that could say no.

1. Task Factory (#92)

loom.com/share/03becd2fe98d44149bd6f274b860f110

by Alazar Shenkute It turns merged GPU speedups into training tasks for coding agents, and times them on real Pixel 5 phones. Then the team hired attackers. Argon agents tried to cheat the grader. Two of four attacks worked: one replayed output tokens, one faked the timer. A fixer closed four holes. Both cheats now score zero.

I like that the demo told us the attacks worked. That is the part that seems like some teams may cut.

Task Factory

Task Factory

2. MigrationProof (#8)

youtube.com/watch?v=1uuKAnFm__4

It ports GPU code to TPUs. The team planted a bug on purpose, a normalization fault. An independent verifier stopped the migration. Argon repaired it, and all 18 contract checks passed. [VERIFY] The demo also says the final hardware handover is still a preview. One honest line, and I trusted the rest more.

by Alpesh Chokshi MS, MBA

MigrationProof

MigrationProof

3. fallguy (#22)

cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/22

by Sudhersan Kunnavakkam Vinchimoor & Ashan Panduwawala

You upload a photo of a real place. It becomes a physics world, and a simulated humanoid takes the fall. Argon plays assistant stunt director. It reads head acceleration, HIC15 and joint torque, then proposes a safer take. The first takes looked fine on screen and would have hurt a person. Apparently the physics said so.

fallguy

fallguy

4. Hindsight (#21)

youtube.com/watch?v=ZgMhqIWLpNs

Robot fleets record thousands of hours of video, but real failures are rare and no one has time to find them. Gemini 4 Argon finds them, rebuilds each one as a digital twin, writes the reward, trains a fix with RL, catches reward hacking, and approves the fix only after checking every site.

by Mahesh K.

Hindsight - Federated Robot Learning

Hindsight - Federated Robot Learning

5. Argus (#41)

youtube.com/watch?v=s9aMt3HxWhk

A San Francisco earthquake, drone video, and noisy resident reports. Argon never sees the damage model’s prediction. Scored against hidden ground truth, its macro-F1 on buildings it saw went from 24 to 46 percent. Across the downtown core, where most buildings never appeared on camera, it went from 33 to 43. [VERIFY] Every damage call points to a second of video.

Argus

Argus

6. FurnishAR (#94)

youtube.com/watch?v=-cS9YEIRdHU

FurnishAR lets you furnish an apartment before you move in. Enter your room dimensions, pick a design style and budget, and a 3-agent Gemini 4 Argon pipeline designs your space with real IKEA products — rendered in an interactive 3D room you can explore. Agent 1 analyzes the room layout, Agent 2 places furniture using Argon’s reasoning across the full product catalog, and Agent 3 runs deterministic collision detection, walkway clearance, and budget verification. If verification fails, the pipeline auto-corrects and re-verifies. Every item links directly to IKEA for purchase. Built with Next.js, Three.js, and Gemini 4 Argon Managed Agents.

by Akshat Patel

FurnishAR

FurnishAR

7. Footage-to-World (#68)

cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/68

Drone video in, Three.js scene out. Then it renders the scene from the original angle, scores the similarity, and rebuilds. Three rounds: 58, 64, 45 percent. [VERIFY] The last round got worse, and the team printed it anyway. Their page calls the result “an honest simulated approximation, not a photoreal capture.” I wish more project pages read that way.

by Christopher Donaldson

Footage to World

Footage to World

8. My Little Brain (#66)

youtube.com/watch?v=c7H80NzR3Tg

A brainy AI you chat with. It reads your live EEG from an OpenBCI headset in Chrome while a 3D brain with eyeballs reacts as you talk. Gemini studies the spectrograms, forms hypotheses, and designs quantum circuits to test them on IonQ, IBM, qBraid and Extropic. It found hidden brain states using 48 of 256 simulations. The team’s own caveat: “model hypotheses, not diagnoses.” Ambitious, and honest about its limits.

By Jacob Valdez Rohit K. Hanah Rahman

My Little Brain

My Little Brain

9. ArgonPulse (#112)

youtube.com/watch?v=fsPl0DP8x9g

A crypto execution engine with ClickHouse (chdb) running inside the process. The database does the arithmetic so the model does not have to. [VERIFY: the team and architecture details are thin. Watch the video or cut.] It is one of only two payments projects in the gallery. An agent that spends money is the hardest version of “the checker decides,” and almost nobody built it.

by Mina Mukherjee

What I took from them

Give every generated output a checker that cannot be argued with. A test suite, a simulator, a fuzzer, a ledger.

Hire an attacker before you ship. If available : ))

Print the score that went down. No spin.

Shack 15 - San Francisco

Shack 15 - San Francisco

My own project, Vaulted Checkout, tried the same idea for agent spending. A vaulted key signs only inside fixed rules, and the model can ask but never sign. The settlement and the refusal were real testnet outcomes. The replay used synthetic timing, and no live Argon purchase happened. The judges thought logs were a thin use of Argon. I agree, and I know what I would do next.

sandboxlabs.ai/vaulted-checkout-explainer

Thank you

Thank you again to team Cerebral Valley & judges: Ivan Leo Joana Carrasqueira Shudhatma Jain subhasmita sahoo Oskar Yao, CFA


Appendix

Winners

Congratulations to the winners!

🥇 First Place: polar bears youtube.com/watch?v=dGreES9C6a0

🥈 Second Place: Morf youtube.com/watch?v=H6O_7u5j_jI

🥉 Third Place: eagle. cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery/75

🎖️ Best Use of Managed Agents: Revive youtube.com/watch?v=pmEjzcfTeiY

🎖️ Honourable Mention: Aniket’s Team (Skywake) skywake.vercel.app

*Part 1 said 136 projects which is what I recall the Hackathon emcee said. The gallery says 128. My guess is that the gallery lists only teams that finished a submission.

Gallery: cerebralvalley.ai/e/gemini-4-argon-sf/hackathon/gallery

bsky.app/profile/schwentker.sandboxlabs.ai/post/3myvo52nw6k2n

twitter.com/schwentker/status/2115263757337452871

← Field Notes