Ayla AI · 2024 – now
Keeping hallucinations out of clinical summaries
Problem
VisitAssist turns patient visits into medical summaries. The summarizer was one 1,333-line class. Any change risked wrong output reaching a doctor.
What I built
A shared LLM pipeline across 139 files, proven with golden-fixture parity tests. An ICD-10 and UMLS validator that checks every summary.
Result
The rewrite shipped with no clinical regressions. The main file is 54% smaller, split into eight modules.



