AI-assisted reimbursement review that turns human decisions into tested, reusable checks.
🏆 HackMIT 2026 Finalist
Finance teams spend time reconciling receipts, chasing missing context, and resolving the same exceptions again. Sift brings the claim, its evidence, and the checks into one workspace: approve supported claims, investigate ambiguity, and put the remaining decisions in front of a reviewer.
Built at HackMIT 2026, motivated by Maximor’s finance-workflow use case.
| What Sift does | |
|---|---|
| 📥 Bring in the paperwork | Upload receipts and supporting documents; inspect extracted fields and suggested claim links before confirming them. |
| 🔎 Check the evidence | Validate amounts, currencies, dates, policy caps, identity, merchant names, and possible duplicates. |
| 🧭 Investigate exceptions | When useful supporting evidence exists, run a bounded investigation with recorded read tools and source citations. |
| 👤 Keep reviewers in control | Show the decisive facts, preserve human decisions, and prepare applicant notifications separately from private review notes. |
| 🧠 Keep the lesson | Turn a supported review reason into a scoped check, test it, and reuse it only when another claim supplies its own matching evidence. |
For example, a hotel receipt may use a billing descriptor that differs from the booking’s hotel name. Sift can connect the documents through their booking reference, guest, date, amount, and currency. A supported reviewer approval can teach that relationship without bypassing overclaim or duplicate checks.
Current app, captured from the isolated 14-claim local showcase with simulated providers.
See the reimbursement workspace
flowchart LR
A[Receipts + supporting documents] --> B[Extract structured evidence]
B --> C[Code checks + Jev assessment]
C --> D[Eligible automatic approval]
C --> E[Investigate unclear evidence]
E --> C
C --> F[Human review]
F --> G[Save decision + held notification]
D --> G
F --> H[Scoped check → safety tests]
H --> I[Activate + recheck pending claims]
I --> C
One Next.js application handles the UI, API routes, and review services. Code enforces financial controls; Jev handles bounded semantic judgments and search; Azure OpenAI powers the investigation planner. Live mode stores records and private originals in Supabase. The local showcase uses private files and simulated providers.
Learning means tested evidence-check reuse, not model retraining. Saved human decisions survive rechecks. Approval records a decision; Sift does not move money.
| Layer | Tools |
|---|---|
| App | |
| UI | |
| AI | |
| Data & email | |
| Testing | |
| Built with |
Devin contributed development, testing, and review support. The investigator inside Sift runs in the app; Devin is a development collaborator. The earlier browser/voice prototype uses Grok voice and lives separately from the reimbursement app.
Requires Node.js 22.18+ and npm. No API keys are needed for the local showcase.
git clone https://github.com/Cryplo/HackMIT26.git
cd HackMIT26/reconciliation
npm ci
NEXT_DIST_DIR=.next-showcase npm run demo -- --showcase --audit-ready --port 3002Open 127.0.0.1:3002/overview and choose Start audit. The launcher creates a fresh private dataset with 14 fictional claims and simulated providers. No real email is sent. Reset demo lets you replay the flow.
The cases cover ordinary claims, amount mismatches, policy-cap violations, duplicate and lookalike purchases, and matching, conflicting, or missing booking evidence. See the showcase guide for the separate human-feedback learning rehearsal and live setup for provider configuration.
One focused offline check:
node --conditions=react-server --import tsx scripts/check-showcase.ts