From a notebook to production medical AI.

The expensive part was not one prediction. It was the whole workflow.

Medical expert witnesses and legal consultants work through large, inconsistent records before they can audit a bill or defend a report. The job crosses documents, billing codes, edits, fee benchmarks, and professional judgment. Automating only one step still leaves the user stitching the rest together.

Docula was built as the connective system: ingest the source material, turn it into structured work, apply domain rules, and return something a professional could inspect and use.

Production quality required a measurement system.

My focus was moving the core workflow from a Gemini notebook into a reliable product. That meant obtaining golden data, breaking the broad task into observable subproblems, incorporating domain-expert feedback, and making failures legible enough to fix.

A demo answers “can this work once?” Evals answer “is the system getting better?”

The useful unit was not a single model score. It was a chain of checks around the workflow—enough context to know where quality moved and why.

This engineering journey became my first public talk, at Vancouver.dev’s event on agentic AI evaluation and benchmarks.