Research

Every workflow we automate, measured in the open.

Each note takes one part of the revenue cycle, runs it on public data against the published baselines, and reports the numbers, the method, and the limits. What holds up is what we ship.

·10 min read

The Convention That Beat Fine-Tuning at ICD-10 Coding

On a public outpatient benchmark with two-coder ICD-10-CM labels, the best open classifier scores micro-F1 0.34. A system that trains nothing on coding data scores 0.74, above every published result on the same notes. Most of the gain came from one convention, confirmed on a held-out split: code what the visit addressed, not everything the note documents.

Read the note →