RegMatch Prototype

Research and Discovery

Planning the MVP

Medicare Advantage audit teams receive more than 300 regulatory requirement documents each year. Reviewers must interpret each requirement, search a large policy library for supporting evidence, and document a traceable response for the final audit.

Reg Match turns that work into a structured workflow: it extracts requirements from the source document, surfaces candidate policy evidence, and gives reviewers a source-linked workspace to inspect and confirm every match.

Challenge and Solution

Turning a PDF into a review workspace and evidence search, designed for transparency

The first prototype focuses on a real hospice review form: a 14-page document containing 64 review findings, evaluated against a library of 373 policy PDFs.

Reg Match reads the uploaded PDF locally, then asks Claude to identify its explicit review requirements as validated structured data: numbers, titles, source references, and categories. For text-rich documents, it sends extracted text to reduce cost and latency; for scanned PDFs, it falls back to the original file. I used a Server-Sent Events stream to report live progress while Claude works - showing the number of requirements found so far - but wait for the complete, validated response before starting evidence search.

I pre-indexed the policy library by page. Instead of re-reading all 373 PDFs every time a reviewer evaluates a requirement, Reg Match searches prepared policy text, filters out common terms, and ranks likely matches by keyword overlap.

Indexing made the workflow both faster and more traceable: reviewers get likely evidence quickly, then inspect its exact source in context. They can also search the library manually or upload an additional policy when the initial search does not surface useful evidence.

RegMatch converts an uploaded regulatory document into structured requirements. Each one receives a short title, review text, source reference, and category label.

Categories use the document's own section headings when available; otherwise, the model creates concise topic-based groupings. The interface preserves that order, transforming a long PDF into a navigable review outline instead of a flat checklist.

Reg Match uses lexical search to surface candidate policies, then selects the passage where the most requirement terms appear together. It expands the excerpt to complete sentences and maps it back to the relevant PDF page. Each recommendation includes a readable excerpt, source policy and page number, in-context PDF viewer, draft status, and explicit confirmation control.

AI can label evidence as supported, needs review, or not found - but the reviewer makes the final decision. The goal is to reduce the work of finding evidence without obscuring where it came from.

Results and Learnings

From MVP to a faster system

After the workflow was working end to end, I focused on its bottlenecks.

I moved standard PDF text extraction to run locally, then sent extracted text to Claude rather than repeatedly asking the model to process an entire document. For scanned or unreliable PDFs, the system falls back to the original PDF to preserve accuracy.

I also restructured evidence retrieval. Instead of generating snippets and page citations for every policy with any keyword overlap, the system ranks the full library first, then creates detailed evidence only for the top results.

On the 64-requirement sample, that reduced evidence-search time from roughly 7 seconds to 0.3 seconds - about a 23× improvement - without changing the returned matches.

AI-native products are not just chat interfaces layered onto an existing workflow. The meaningful work was deciding where automation could accelerate review, where the system needed to show its reasoning, and where a person needed to stay in control.

Reg Match is a prototype, but it demonstrates a practical direction for AI-assisted policy review: faster document processing, traceable evidence, and deliberate human confirmation before a decision is final.

View the Github repo and live prototype.