A confusing letter arrived. One missed date could matter.

Benefit renewal notices can bury dates, checkboxes, and conditional documents inside dense text. Fineprint helps a reader find what the letter actually says and check it against the source.

What it does

Fineprint is a private, browser-only evidence reader for renewal notices. Paste text or drop a text-based PDF. It organizes selected timing excerpts, program check-marks, document conditions, renewal steps, and contacts. Every item links back to its exact words and page in the original.

Dates appear as verbatim excerpts in source order. Fineprint does not decide which date is your deadline or whether you are eligible. Conditional documents retain their conditions. Unreadable check-marks are labelled unreadable, scanned PDFs get a clear no-OCR warning, and unsupported documents do not receive a confident checklist. Nothing is uploaded and no AI service is called at runtime.

How it was built

A static Vite + strict TypeScript app uses pdf.js for PDF text and vector checkbox extraction. Deterministic rules create an evidence model whose quotes are checked word-for-word against source lines. Sentence-aware extraction preserves context across the tested soft-wrapped negation cases. GitHub Actions runs lint, typecheck, unit and corpus tests, browser E2E, and the production build.

The hard lesson

Early versions tried to infer deadlines. Independent review found historical dates mistaken for current instructions and negation clipped at line breaks. Local phrase fixes kept moving the problem. The redesign removed automatic deadline conclusions from the type system and replaced them with a quote-first evidence index. The reader remains responsible for checking the complete letter.

Evidence, without overstating it

The regression suite has 90 unit tests, a 69-case author-built corpus with 201 expected findings, and 14 browser E2E tests. These cover citation integrity, abstention, annotation leaks, mobile layouts, 200% zoom, file rejection, and reset behavior. This corpus was visible during development: it is regression coverage, not held-out or population accuracy. See the repository's QA and repair reports for the failure history and fixes.

Limitations

Novel layouts can produce incomplete excerpts. No OCR, eligibility, medical, or legal determinations are provided. Users must check the full notice. Known limitations include contact text inside some document excerpts, a single highlighted source line per click, and possible omission of real lines beginning “Source:” by annotation filtering. No validation on real, filled personal notices is claimed.

Attribution and AI assistance

Built for Sharon Basovich with AI-led assistance from dot and Devin (Cognition), including planning, implementation, testing, review coordination, documentation, and demo preparation. Sharon authorized the project and campaign; this does not imply that Sharon personally wrote or reviewed the code. The narration uses ElevenLabs text-to-speech. The app itself is deterministic TypeScript with no runtime AI.

The historical blank Texas HHSC H1830-R SAMPLE form (December 2018) is credited in SOURCE-MANIFEST.md. Only extracted text is bundled, not the PDF; a redistribution license has not been confirmed. It is a historical document-comprehension example, not current policy guidance. Other demo letters are original, labelled synthetic examples.

Try the demo

Open the live app and choose “Try it with the sample letter.” Inspect “Dates & timing mentioned,” jump to an exact source line, then inspect the full conditional housing-cost evidence. Try a utility bill to see the unsupported-document state.

Live app: https://sharonbasovich.github.io/fineprint-warriorhacks/ Source and QA reports: https://github.com/sharonbasovich/fineprint-warriorhacks Direct video fallback: https://github.com/sharonbasovich/fineprint-warriorhacks/blob/main/docs/video/fineprint-walkthrough.mp4

Built With

Share this project:

Updates

Submission history