03 / AI agents / Insurance
Coverage Brain
Four AI agents that give an insurance broker one memory of every client conversation.
Live app / the walkthrough
01 / The problem
Clients mention changes. Nobody logs them.
Clients mention changes to their brokers all the time: two new vans, a son who just got his licence, a second office, drivers starting next month. Each one can open a gap between what the client now does and what the policy covers, and most of them arrive as good news in a passing email, not as a request.
02 / How it works
Four agents, one job each.
Each agent does one job, hands a single packet to the next, and stops the chain if something is missing.
| Agent | What it does |
|---|---|
| Watcher | Notices what changed. Reads the emails, call notes and documents for one case, and pulls out every change with a date and the exact words it came from. |
| Matcher | Checks the policy on file, and flags an exposure with nothing covering it as a potential gap for the broker to review. |
| Closer | Drafts the follow-up in plain language, tied to the client’s own words. It stays unsent until the broker approves it. |
| Scribe | Keeps the record: what changed, what was offered and when, all timestamped, so it protects the agency if a claim is ever disputed. |
03 / The guardrails
Nothing is sent without the broker.
The agents never guess at coverage, never send anything and never edit records on their own. If a source is missing, or a fact can’t be traced back to the original text, the chain stops and the broker decides.
04 / My part
The demand, the purpose and the memory.
My part: led the market research into broker demand, aligned the team on the app’s goals and purpose, and installed the memory layer in Cognee.
05 / The result
First place at the Personal Agent Hackathon, NYU Tandon.
Coverage Brain won first place at the Personal Agent Hackathon, NYU Tandon, where it was built as a team project: four agents that keep one memory of every client conversation, with the broker approving everything that goes out.