Coverage Compass

Preview

Inspiration

Life insurance is one of the most important financial decisions a family makes, and one of the most confusing. Most "needs calculators" are long forms that spit out a number with no explanation. We wanted something that feels like a conversation and shows its work, so people can trust the number.

What it does

Coverage Compass is a conversational life insurance needs analyzer. It asks questions one at a time and adapts as you answer: no kids means no college-cost questions, and renting means no mortgage-term question. It then shows how much coverage might fit your family and every line of the math behind it. You can describe your situation in plain language, and the app turns it into structured inputs.

How we built it

The AI never does the math. The coverage amount comes from a tested, deterministic calculation that the server computes. Claude only parses free-text answers and explains the results.
Every figure is verified. A validator extracts each dollar amount in Claude's response and checks it against the server-computed numbers (within 5%). If anything doesn't match, the response is withheld and a standard explanation is shown instead.
Graceful fallbacks. If the model is slow or unavailable, a built-in parser and standard answers keep the user's progress intact.
Secure by design. Argon2id password hashing, Google sign-in, a strict content security policy, and API keys kept server-side in AWS Secrets Manager.
Production-style delivery. The backend is a FastAPI service with PostgreSQL on AWS RDS, deployed on ECS, with the React frontend on Amplify. CI/CD runs through GitHub Actions and blocks deploys that contain unapplied database migrations.

Challenges we ran into

Making an LLM trustworthy in a financial context. We solved it by treating the model as untrusted and verifying its output instead of relying on prompt instructions alone.
Keeping the conversation natural while still collecting exactly the inputs the calculation needs.
Shipping a deployed, authenticated, tested full-stack app under a hackathon deadline.

Accomplishments that we're proud of

A verification layer that checks every dollar figure the AI writes before a user sees it
More than 500 automated tests (339 backend, 169 frontend), with about 98% backend coverage
A live deployment with real authentication and CI/CD, not just a demo

What we learned

Good AI features come from clear boundaries: let deterministic code do what it's good at, let the model do what it's good at, and verify the handoff. We also learned how much a full-stack build under a deadline depends on teamwork and a clear split of responsibilities.

What's next for Coverage Compass

Broader coverage-type guidance, more checks on the AI's output beyond dollar figures, and more accessibility and localization work.

Built With

python · fastapi · react · postgresql · amazon-web-services · amazon-ecs · amazon-rds · aws-amplify · aws-secrets-manager · claude-api · anthropic · github-actions · argon2 · google-oauth