FinAssist - Payment Resolution Agent

An AI agent that resolves failed and unsettled payments by giving users honest, policy-accurate answers when their money is stuck.

Built with a four-loop architecture: agent loop, verification loop, event-driven lifecycle, and a human-gated self-improvement loop.

What it does

• Classifies payment issues and retrieves transaction data deterministically
• Composes regulation-accurate responses grounded in NPCI and NEFT policy rules
• Blocks unsafe responses via a deterministic policy layer before they reach the user
• Tracks tickets until the payment is resolved, not just until the user gets an answer
• Learns from feedback through a nightly eval pipeline where a stronger model grades output and humans approve changes

Stack

FastAPI, Supabase, Anthropic API, OpenTelemetry + Arize

Setup

git clone https://github.com/AIPMProfile/payment-resolution-agent
cd payment-resolution-agent
pip install -r requirements.txt
cp .env.example .env
python -m app.db.seed
uvicorn app.main:app --reload

Required keys in .env:

ANTHROPIC_API_KEY=
SUPABASE_URL=
SUPABASE_KEY=
ADMIN_API_KEY=
AUTH_SECRET_KEY=
ARIZE_API_KEY=
ARIZE_SPACE_ID=

Run app/db/supabase_schema.sql in Supabase before seeding.

Chat: http://localhost:8000
Admin: http://localhost:8000/admin/?key=YOUR_ADMIN_KEY

Tests

pytest tests/ -v

Known gaps

Transactions table is seeded and static. In production, settled_at would be populated by Federal Bank webhook on NPCI settlement confirmation — the cron and ticket lifecycle are built for this integration.