Clinical AI

At Abridge I build AI-generated clinical documentation. What makes this domain distinctive: model capability isn't the bottleneck —trust is. Error tolerance is effectively zero, and the bar isn't "impressive" but "trusted by a physician on a busy day." That shapes how my work splits: half building the product, half building the system that lets the product change safely.

Note generation & the eval system

The hard part of generation (NoteGen) was never producing one good note — it's holding a stable floor across wildly different real-world encounters. So the other half of my effort goes into the evaluation system: LLM-as-judge calibration, golden and challenge sets, promotion gates — in essence, giving the team grounds to say "this change is safe to ship." An eval system isn't a QA checkpoint; it's the product's transmission. It sets how fast we dare to evolve.

The wider surface

Along the same throughline, I keep hunting for structural leverage: consolidating scattered generation logic into composable agent workflows; turning prompt tuning from a craft into an automated process (APO); exploring agent memory so the system accumulates judgment across sessions.