Healthcare Claims Intelligence
- Sector
- Health plans & payers
- My role
- Team lead & developer
- Key decisions
- 4 covered below
In plain terms
Turns piles of messy claim documents into clean, structured summaries, so physicians spend their time deciding instead of reading.
The problem
Physicians reviewing claims were spending more time reading documents than making decisions. The documents arrived in volume, in a dozen formats, with no consistent structure. Sending every page to a model would have been very expensive — and, given what the pages contain, not allowed.
Architecture
Scroll sideways to see the whole diagram.
Key decisions
- 01
A document intelligence layer
Layout-aware parsing, field-level extraction and schema validation turn raw records into fields a physician can scan in seconds.
- 02
Redaction the client could inspect
Sensitive fields are detected and tokenized before anything reaches a reasoning model. We built this in-house rather than buying it, so the client could audit it line by line instead of trusting a vendor.
- 03
Queue first, not request-response
Redis-backed background jobs let large batches queue and drain predictably instead of tying up a request.
- 04
Semantic caching
Similar extraction queries across similar claims are answered from cache instead of calling the model again — a direct, measurable cut in processing cost.
Built with
- TypeScript
- Claude API
- PostgreSQL
- Redis
- In-house PHI layer