Taccid

Healthcare admin workers spend hours on repetitive paperwork like prior authorization instead of focusing on patients. AI agents can take that on, but there was no good way to test or train them on these workflows. That's what we built: simulated healthcare environments and testing infrastructure that frontier labs could use to benchmark and improve their agents.

These tasks had verifiable answers. "Did the patient try these 5 medications?" "What's the exact primary insurance code?" Outputs could be graded like unit tests, not essays, which made reinforcement learning possible.

The system had four pieces: a simulated insurance portal (React/Next.js frontend, Python backend, synthetic patient data from Synthea), a task engine where each task is a JSON spec with patient, drug, insurer, and expected outputs scored by Python grader functions, Daytona for spinning up thousands of isolated environments in parallel, and Harbor for standardized packaging so any lab could plug our environments into their training stack.