The problem
Documentation burden is the single biggest driver of clinician burnout. Existing scribing tools generated text — but not the structured note clinicians could sign. Citations to the source utterance were essentially nonexistent.
Our approach
We built a scribe pipeline with two non-negotiable invariants:
- Every claim has a source. The clinician can hover any sentence and see the ambient transcript span that produced it.
- Templates are owned by the clinical team. Note structure (SOAP, narrative, etc.) is defined per-specialty in the CMS; the agent fills the slots.
The first version we shipped had a 12% citation-failure rate. The third had 0.8%. The evals caught every regression.— Eng lead, Healthtech, US
The architecture
What we actually ship.
Every system we build follows this shape. Client at the edge, tools in a sandbox, traces everywhere, evaluators gating output. No black boxes, no "it works on my machine."
A run, on this system
Watch an agent do the job.
Three real production scenarios, replayed at observed latency. Every box is a span; every span has tokens, cost, and an eval gate. This is what shows up in your traces, not a marketing animation.
Scribe pipeline: ASR with medical lexicon, retrieve patient context + template, draft per-SOAP-section, evaluator gates clinical-safety claims.