AI-assisted reimbursement review
Our team built Sift to help finance teams work through reimbursement claims, investigate unclear receipts, and carry useful lessons from one review into the next. Inspired by Maximor's finance-workflow use case, it brings receipts, supporting documents, automated checks, and human decisions into one workspace.
The problem
A receipt does not always tell the whole story. A hotel charge might appear under a booking company's name, or a claim might need another document to explain a mismatch. Reviewers need a way to separate valid exceptions from unsupported claims without repeating the same investigation every time.
What we built
Sift extracts information from uploaded paperwork and checks it against the claim and reimbursement policy. Supported claims can qualify for automatic approval. When useful supporting evidence is available, a bounded AI investigation examines it and records the sources behind its findings. Unresolved cases go to a reviewer with the relevant evidence in view.
The feedback loop is the part that makes Sift interesting. A supported human approval can produce a narrowly scoped check that must pass safety tests before it is activated. Future claims still need their own corroborating documents, and saved human decisions remain intact when checks run again.
One example: two names, one hotel stay
A receipt says “Harbor Reservations,” while the booking confirmation says “Harbor Hotel.” Matching booking references, guest details, dates, amounts, and currencies can establish the relationship.
If a reviewer approves the claim for that supported reason, Sift can save the relationship as a tested check. Another claim can use it only when its own evidence supports the same match. Amount, policy-limit, and duplicate checks still apply.
How it works
We combined explicit financial checks with AI for questions that require interpreting documents. Application code checks amounts, currencies, dates, and policy caps. Jev handles bounded semantic assessments and search; Azure OpenAI powers the investigation planner. The app records investigation steps and citations so reviewers can inspect the evidence.
The review-learning path currently supports the hotel billing-descriptor case above. It saves reusable evidence-check logic rather than retraining a model, and it excludes one-time exceptions and proposed policy changes from automatic rule creation.
Built with: Next.js, React, TypeScript, Tailwind CSS, shadcn/ui, Supabase/PostgreSQL, Jev, and Azure OpenAI. Verification tooling includes the Node test runner, Playwright, and PGlite.
Explore the project
View the source and local demo instructions
Sift is a hackathon prototype using synthetic claims. The local showcase runs with simulated providers; approval records a decision and does not transfer money.