I’m the developer of NoteCoder, an experimental browser-local tool for reviewing diagnosis wording and suggested ICD-10-CM codes. I’d like feedback on the separation between candidate discovery, model decisions, and what a reviewer can actually verify in the source text.
The public pipeline first proposes candidates from dictionary phrases, associated diagnosis wording, and details such as stage or laterality. A local decision model then evaluates diagnosis and qualifier questions against the complete note. This creates two distinct failure points: a relevant candidate can be missed before inference, or a proposed candidate can receive an incorrect model decision.
The production app runs inference with WebAssembly. The default model’s initial download is about 2 GB; browser support, peak memory and latency vary by device. Notes, prompts and results stay in tab memory. Asset/model downloads still make network requests, and the browser may cache or evict model files. I’m not claiming measured browser performance from separate native-model experiments.
A reproducible synthetic walkthrough:
The usage guide documents this flow and the effect of changing scope, model or prompt. The Medicare V28 and ACA HHS 2026 options are mapping filters, not complete diagnosis coverage or reimbursement determinations.
For evaluation, would you separate missed-candidate recall from assertion/qualifier errors, and how would you test whether an interface makes superseded wording look falsely authoritative? Browser memory/ observations would also be useful. Please keep shared examples synthetic. The app is free, has no account requirement, and has optional project support. It can miss diagnoses and suggest incorrect codes; human review is required. It is not clinically validated or a billing system. AI assisted development and preparation of this post.