Pre-Dispute Deflection Shield — a refund-or-contest engine for disputed payments. When a chargeback alert arrives, it decides whether to refund outright or fight it. The counterintuitive part: sometimes you refund a dispute you'd win, because card networks penalise you on dispute ratio, so the slot costs more than the money. It beats the strongest baseline by ~₹227k per 1000 alerts, contesting 60 of 600 at a 75% win rate against a 48.8% base. The LLM stays off the money path by design — complaint text is adversarial, so its influence on net is 0.33%, checksum-verified. Python, Scikit-learn, Razorpay API, Streamlit.

– Built a standalone MCP server exposing a search_hotels tool over stdio for live availability and pricing. – Bounded the agent at two tool rounds; returns empty results honestly rather than inventing hotels that don't exist. – Gemini agent grounded in live DB context, resolving named places within a 1.5 km radius; JWT auth with enumeration-safe login