Inspiration

Fixing an infrastructure incident and getting the service credit owed under a contract are two different jobs.

Even after service is restored, someone still has to check the contract, find the timestamps, collect evidence, and follow up on the claim. InfraClaim brings those steps into one place.

What it does

InfraClaim helps an operator work through a cooling SLA claim.

The operator adds the contract terms, records what happened, and attaches supporting evidence. The app shows the proposed credit, the evidence behind it, and anything that needs a closer look. A separate reviewer can approve the case, reject it, or ask for more information.

The demo also follows a claim through a simulated vendor dispute and acceptance, then checks it against a simulated credit memo. Acceptance alone does not count as recovered money.

Cases and their history stay saved when the app restarts.

How we built it

The frontend uses Next.js, React and TypeScript. The backend uses Python and FastAPI, with SQLite for storage.

NVIDIA Nemotron runs through Nebius Token Factory. It assesses the evidence and returns a structured response with references to the supplied records.

The model does not decide the credit amount by itself. Code calculates the amount and applies the limits. Missing evidence, conflicting records or an invalid model response can hold the case for review.

Operator, reviewer and administrator accounts have different permissions. Provider keys stay on the backend.

Challenges we ran into

The difficult part was deciding when the app should stop.

An answer can sound convincing even when the evidence is incomplete. We added checks for citations, contradictions, missing information and invalid responses, along with a separate review step.

Security testing also found problems that the original tests had missed, including alternate routes that bypassed permissions. Those findings led to fixes and additional regression tests.

Accomplishments we're proud of

The current version passes 129 backend tests and a production frontend build.

We also tested a complete simulated claim, checked that cases survive a restart, and confirmed that repeating reconciliation does not count the same credit twice.

A recorded live evaluation completed 50 successful Nemotron calls across 50 synthetic cases. All produced valid assessments and the expected guarded outcomes in that test set. The model received calculated context, so these results are not a claim of accuracy on real customer cases.

What we learned

Getting a model response was only one part of the work. The application also needed to handle uncertainty, preserve evidence, enforce permissions and recover from failures.

What's next

The next step is testing with real customer cases and connecting incident, vendor and credit systems.

For now, InfraClaim is an internal pilot with synthetic examples. Vendor responses and credits are simulated. We have not recovered real money or validated the product with customers.

Built With

Share this project:

Updates

Submission history