AI intelligence layer for Bank Reconciliation. FastBank Hackathon — May 26–28, 2026.
A standalone React app that imitates the FastBank visual language and adds four AI-powered features on top of a reconciliation workflow:
- Close Guarantee — opens every session with a probability score and known blockers
- Match Debugger / TRACE — explains why an unmatched transaction didn't match
- Risk Firewall — inline alerts while the user is resolving items
- Reconciliation Narrator — generates a full audit document at close
All four are powered by a single Financial Brain context (data/financial-brain.json) that is injected into every AI call. That's what makes the demo feel like the system "remembers" the customer.
The data is mocked. The AI reasoning is real (Claude API).
# Use the right Node version
nvm use
# Set your API key
cp .env.example .env
# edit .env, paste your ANTHROPIC_API_KEY
# Install everything
npm run install:all
# Run both client and server with one command
npm run dev- Client: http://localhost:5173
- Server: http://localhost:4000
- Health check: http://localhost:5173/api/health (proxied)
client/ Vite + React + TypeScript + Tailwind frontend
server/ Express + TypeScript backend (Claude API proxy + Brain context)
data/ Mock JSON files (Financial Brain, bank/SAP transactions, etc.)
CLAUDE.md Architecture rules + conventions. Read this before coding.
CLAUDE.md in this repo root. It defines the architecture, the JSON contracts, the fallback strategy, and the per-feature pattern. Both team members and any AI coding assistant should follow it.
The server has a DEMO_MODE=true env flag that forces all AI endpoints to return mock data. Use it during the live demo if the venue WiFi is shaky. Every endpoint also falls back to a mock automatically if the live Claude call fails or returns invalid JSON — the demo cannot break from a network issue.