The problem
Collections is repetitive, rule-heavy work: a customer says they’ll pay next week, or that they were double-charged, or asks for an instalment plan. Each answer depends on company policy, and each outcome has to land in the payment system correctly.
What I built
Resolve.ai is a FastAPI backend where a LangGraph agent graph handles a collections conversation:
- Classify intent — paying now, promising to pay, disputing, asking for a plan.
- Apply policy — the agent retrieves the relevant policy sections from a Pinecone index and is only allowed to offer what policy permits.
- Change real state — create a Razorpay payment link, record a promise-to-pay, or open a dispute.
Agents here don’t just reply; they use tools and change backend state, so every step is persisted in PostgreSQL and can be replayed.
Reconciliation
The agent never decides that an invoice is paid. Razorpay webhooks do. When a payment lands, the webhook reconciles it against the open invoice and the client sees the status change live over SSE.
What’s next
- Move from Vercel’s subdomain to
resolve.vansshagarrwal.in, with a demo tenant anyone can try. - Put it in front of a few small businesses that chase payments over WhatsApp today.