Last Saturday, I went to the OpenAI Hackathon in SF, hosted at Entrepreneurs First with Think Technologies and The AI Collective. I originally just went in to mess around with the new Workspace Agents and see what cool people I'd meet, but I ended up getting a lot more out of the day than I expected.
The day started with a keynote from Ben Tyson, head of SMB channel partnerships at OpenAI. He started his talk with the Roger Bannister story which is the idea that before 1954, people believed running a sub-4-minute mile was physically impossible, and that once Bannister did it, dozens of other runners also did it in the next year. (Tyson told this as motivational framing; historians have poked holes in the "doctors believed it would kill you" version, but the point about shifting baselines was expressed regardless.) His argument was that we're at a similar threshold with AI: the ceiling for what small businesses think is possible is still set way too low, and Workspace Agents are trying to close that gap by moving from single-prompt interactions to multi-step workflows that run autonomously in the background.
Anyway, back to the actual hackathon. I went by myself, but I ended up meeting so many cool builders and founders. Amongst these people, I met 3 girls. One of them used to be a neurotech consultant, and we started talking about cognitive bandwidth and the specific friction she'd seen in clinics with practitioners dealing with asynchronous data streams, managing patient intakes, synthesizing complex histories, trying not to burn out. We, and the other 2 girls we had just met, felt like this repetitive and high-friction problem could highly benefit from an agentic workflow.
So, we teamed up and built Beacon!
Beacon is a tool designed to run in the background, aggregate asynchronous inputs, cross-reference them, and prep the data layer before a human steps in. Since we were building for the technical track, we used Codex and the Agent SDK to build the integrations and get to a production-quality build.
The coolest part is that none of us knew each other before that morning. Building software with strangers in 4 hours is something I don't get to do very often, and it ended up being a good demonstration of Conway's Law: the adage that software structures inevitably mirror the communication structures of the teams that build them. Because we had to divide and conquer instantly, our architecture naturally evolved into a modular multi-agent pipeline.
I also had to learn a lot about OR logistics really fast. Day-of-surgery cancellations are a big, expensive problem and a single missing lab result (like an expired cardiac clearance or a missing consent form) can derail an entire surgical schedule. To solve this, we split Beacon into four agents that act like a digital relay team:
- Schedule Monitor — a cron job that watches tomorrow's date and pulls all upcoming OR cases.
- Readiness Reviewer — the analytical engine. It ingests the Case ID and patient data and calculates a Readiness Score from 0–100, flagging cases from "Ready" to "At Risk" or "Blocked" the moment something's missing.
- Care Coordinator — the executive function layer. It takes the readiness output and turns any gaps into concrete actions: targeted notifications, escalations, open tickets.
- Briefing Generator — synthesizes the entire web of case data into a concise pre-operative briefing for the surgical team.
It was a chaotic 4 hours, but we got a working build and a demo done right before the 4 PM deadline. And we ended up winning the technical track!
You can check out the source code on my GitHub here, and play with the live Vercel deployment here.
Overall, it was really cool to see what other people built at the hackathon and what many of them are building outside the hackathon context too. I can't wait to keep exploring SF and meeting other builders at events like this!
Things I referenced