About Reclaym

Reclaym is an India-first revenue-recovery agent for failed UPI AutoPay and e-mandate charges, built for the Razorpay Buildathon (Track 03) by Amar Mahato. UPI AutoPay mandates fail at 8 to 15 percent in India, against 2 to 3 percent for cards, and that leakage is largely invisible to the dunning tools built around card retries.

Reclaym turns every failed charge into a tracked recovery case. An LLM diagnoses the root cause, a deterministic expected-value engine selects the bounded action that recovers the most money net of cost, a policy engine gates that action against India's NPCI and RBI rules (the 1 plus 3 retry cap, peak-hour execution windows, and AFA amount tiers), and an executor acts on Razorpay test-mode rails with database-enforced idempotency so a timeout or a webhook redelivery can never become a double charge.

The governing principle is that the LLM proposes and a deterministic engine disposes: money never moves on a model's say-so. Every step is written to an append-only audit trail, and recovery is measured honestly against a randomized control arm on a held-out batch, with confidence intervals and a disclosed list of where the system does not beat control.

Reclaym runs entirely in Razorpay test mode. No real money moves. The full source, architecture notes, and a reproducible evaluation are public on GitHub.