Overview
ClaimBridge is my fourth-year capstone project (ECE496) at the University of Toronto, built with a team of four. It is an AI-assisted intake platform for insurance claims: it turns a claimant's spoken or written account into structured data, and puts a claims adjuster in charge of reviewing it before anything is accepted.
The problem
Claim intake is slow and repetitive. A person has to listen to or read an account, pull out the same fields every time, and keep sensitive health information protected along the way. Automating that carelessly is risky, because a wrong or leaked field is worse than a slow one.
Context and motivation
This started as a capstone proposal in October 2026. I am drawn to problems where a model meets a real workflow, and health insurance claims are a good example: the data is sensitive, the stakes are real, and the useful system is the one people can trust.
My role
- Role on the team: to write. Replace this with what you personally own.
- Contributing alongside three teammates.
Approach
The design principle is that a person stays in the loop. The system proposes, and an adjuster approves. Privacy is handled early in the pipeline rather than bolted on at the end.
Technical architecture
The planned pipeline runs in stages:
- Transcription of the claimant's audio with Whisper (large-v3-turbo).
- Redaction of personal and health information with Presidio.
- Entity extraction into structured claim fields with Phi-4-mini-instruct.
- Storage in PostgreSQL.
- Review in an adjuster-facing interface, where each extracted field can be accepted or corrected.
This is the plan, not a finished system. The details below will be updated as the build progresses.
Implementation
To write. Screenshots, schema excerpts and the review interface go here once they exist.
Trade-offs and decisions
| Decision | Trade-off | Chosen |
|---|---|---|
| Run models locally | More control over sensitive data, but heavier to deploy | To confirm |
| Redact before extraction | Safer, but the extractor sees less context | To confirm |
| Mandatory human review | Slower than full automation, but safer | Yes |
Results and impact
In progress. There are no results to report yet, and none will be invented. The interim demo is targeted for 1 December 2026.
Lessons learned
To write once there is something to learn from.
What's next
Interim demo, then evaluation of the extraction step against real claim-style data.
Links and resources
Repository, slides and demo links go here when they are public.