Competition · AI Agents
Speedrun Clinic Agent
An AI agent, shipped as a Docker image, that works a queue of back-office tasks in a synthetic clinic web app. At Speedrun by Scope, a wrong answer cost a point and a skip was worth zero, so the agent skips any task the evidence doesn't settle. It was 5th of 29 finalists going into the hidden final set (924 tasks in 60 minutes, for a synthetic clinic no team had seen) and finished 1st. It scored 291 on that set, where no other finalist scored above 211. I wrote the starting brief that laid out the strategy, and AI coding agents wrote and tested the code under my direction.
What it does
Drives the clinic web app with Python and Playwright
Messages in the queue can set new workflow rules that change later answers; a compiler turns new rules into code at runtime
A 2,112-call benchmark picked a model for each call type (planner, extractor, task parser, rule coder) on accuracy, latency and cost
Tech stack