ClaimTrace: a Multimodal Evidence Intelligence Platform

ClaimTrace is an AI-powered evidence review platform that helps ecommerce teams analyze returns, warranty claims, and disputes.

Comment

Inspiration

Returns fraud costs ecommerce retailers billions annually. Yet most teams still review claims manually — one blurry photo at a time. We asked: what if an AI could review every piece of evidence simultaneously, the way a seasoned investigator would? With Human review in different stages.

What it does

ClaimTrace lets returns and dispute teams upload evidence (photos, receipts, screenshots) against a claim and receive an instant AI verdict — a confidence score, a recommendation (approve / reject / escalate), and a plain-English explanation of exactly why.

How we built it

React frontend, Express.js API, PostgreSQL for claim history, Amazon S3 for evidence storage, and Amazon Bedrock Nova Lite as the multimodal AI brain — analyzing multiple images in a single call and returning structured JSON decisions.

Challenges we ran into

Getting multimodal prompts to return consistent, structured output at scale. Also wiring S3 uploads through the server to avoid browser CORS restrictions, and building a Vercel serverless deployment that correctly bundles a full Express app. Reducing the size of the index page, website, packaging, SDK, to load website quickly.

Accomplishments that we're proud of

A genuinely production-ready AI pipeline: Nova Lite reviews up to a dozen evidence images per claim simultaneously and returns actionable verdicts in seconds. Zero server-side AWS secrets — credentials stay in the user's browser and has the option to erase.

What we learned

Structured output contracts in multimodal prompts are everything. Telling the model exactly what JSON schema to return — confidence, recommendation, flags, explanation — produces far more reliable and actionable results than open-ended prompting. Building a basic production ready webapp end to end.

What's next for ClaimTrace

Cross-claim fraud pattern detection, direct integrations with other platforms, a customizable rules engine on top of the AI verdict, and analyst productivity dashboards tracking approval rates and time-to-decision by product category. Reviewing the whole project for improvements on - features , security , user experience, output confidence. Working on the feedback.

Built With

Share this project:

Updates