The problem

A visit note says right wrist. The operative summary says left. A patient preparing a correction request has to find both passages, identify their sources, and explain what wording they want reviewed. ChartRepair keeps those steps together.

What it does

Open a fictional case or select two text-based PDFs. Scan for possible differences or search in plain language. Inspect each candidate on the highlighted original page, confirm the evidence, and export an editable request as PDF or JSON. Check a follow-up for exact requested wording and read the surrounding text. One fictional follow-up quotes a request while stating that the record has yet to be amended.

How we built it

A bundled MiniLM model runs in a browser worker and ranks related passages. The scan combines embeddings with rules for sides, dates, times, and dose strings. PDF.js maps selected passages back to the source page. The user decides which passages to cite and what wording to request. Codex helped build the browser worker, source highlights, fictional cases, opening animation handoff, and regression checks.

Try it

Extract the source ZIP and run node scripts/serve.mjs. Choose Try a sample case, scan the records, confirm one passage per source, and save a draft. Then check the fictional follow-up. The case picker also includes scheduling and medication documentation examples.

Verification

The build passed 15 unit tests and the Chrome browser suite at three viewport widths. The prototype uses fictional records for testing and supports selectable English text.

Built With

Share this project:

Updates

Submission history