Winning Two Awards at the Intelligence at the Frontier Hackathon

The Intelligence at the Frontier hackathon, hosted by Funding the Commons and Protocol Labs in San Francisco (February 23-24, 2026), was one of those events where the ambition level matched the tooling. Around 150-200 curated builders, a $26,750+ prize pool, and four tracks spanning physical AI, agentic systems, AI safety, and sovereign infrastructure. Our team walked away with two awards.

HydraSwarm: Best Overall Use of DeepLake

The first win came from HydraSwarm — a self-improving multi-agent AI system where 7 specialized agents (Product Manager, Architect, Developer, Reviewer, QA Engineer, SRE, and CTO) collaborate on software engineering tasks.

The key insight was institutional memory. Most AI systems generate once and forget. HydraSwarm writes lessons back to HydraDB (powered by DeepLake) after every run. The next time a similar task comes in, each agent recalls relevant knowledge before generating output. The improvement is measurable: Run 1 scores 7/10, Run 2 scores 8/10, Run 3 scores 9/10.

We built live agent thinking logs showing every HydraDB query and storage operation, SSE streaming for real-time agent activation, and a run comparison view that shows score deltas side-by-side. The judges could see the improvement happening in real time. 326 unit tests across 21 suites gave us confidence to ship fast without breaking things.

Physical AI & Robotics Track: Unitree G1 Pipeline

The second win was in the Physical AI & Robotics track, sponsored by NomadicML. This one was hands-on — literally. We built a rapid-iteration pipeline for the Unitree G1 humanoid robot, moving from manual demonstration to autonomous policy testing.

The stack: Meta Quest 3 integrated with MuJoCo for intuitive teleoperation, NVIDIA Sonic for low-latency control commands, DeepLake for high-throughput tensor storage and fast I/O of demonstration data, and NVIDIA GR00T for fine-tuning locomotion policies. We used Nomadic AI to diagnose failure modes when the robot's walking or pick-and-place tasks went wrong.

The hardest part wasn't any single component — it was getting them all to work together under time pressure. When you're debugging a humanoid robot at 2 AM and the walking policy keeps falling over, you learn to prioritize ruthlessly.

MedAssist: Robotic Agents Hackathon Follow-Up

A few weeks later at the Robotic Agents Hackathon (March 2026), our team built MedAssist — an AI-powered medication verification and dispensing agent. The system closes the full sense-reason-act loop: a camera on a SO-101 robotic arm scans the medication tray, Claude AI reasons about what should be dispensed for a specific patient through a 10-check safety sequence, and the arm physically picks the verified medication. ElevenLabs narrates each action in real time.

This one was personal. Approximately 1.5 million people are harmed by medication errors in the US each year. Building a system that could autonomously verify and pick the right medication — or refuse to move if something is wrong — felt like the kind of work that matters.

What I Took Away

Hackathons at this level aren't about demos. They're about proving you can ship a working system under extreme constraints. The skills that mattered most were systems integration, rapid debugging, and knowing when to simplify. The robotics track especially tested that — you can't fake latency when a physical robot is moving in front of judges.