RECOVER-ALLOC — Don't Just Retry Payments. Optimize Revenue Recovery | Razorpay AI Buildathon 2026

RECOVER-ALLOC — AI Revenue Recovery Allocation Engine Razorpay AI Buildathon 2026 Track: AI Revenue Recovery Solo Build The problem isn't knowing how to retry a failed payment. The problem is deciding where to spend limited recovery capacity. RECOVER-ALLOC is a constrained AI revenue recovery allocation engine that determines which revenue-at-risk accounts should receive which recovery intervention while respecting limited merchant resources and deterministic policy constraints. In this demo, the system considers: • Payment retries • WhatsApp reminders • Human escalation • Limited retry slots • Limited messaging capacity • Limited human support hours • Merchant policy constraints CORE APPROACH I model revenue recovery as a Multiple-Choice Multidimensional Knapsack Problem (MCMKP). For each account-intervention pair, the system estimates expected net recovery and then optimizes the complete portfolio of recovery decisions under competing resource constraints. Rather than simply selecting the highest-value accounts independently, RECOVER-ALLOC reasons about resource competition across the entire recovery batch. OPTIMIZATION VERIFICATION A concrete counterexample demonstrates: Naive Greedy Objective: ₹24,845.50 MCMKP Optimal Objective: ₹27,644.30 Improvement: 11.26% The CP-SAT optimizer was independently verified against a brute-force solver on 100 randomized instances, matching the optimal solution in 100/100 cases. AI + DETERMINISTIC CONTROL The system separates AI prediction from execution authority: Evidence → Diagnosis → Probability → MCMKP Optimization → Policy → Execution The LLM is restricted to diagnostic assistance and does not receive direct execution authority. Deterministic policy gates control whether proposed interventions can actually be dispatched. RELIABILITY & SAFETY RECOVER-ALLOC includes: • Idempotent execution • Duplicate-action protection • Explicit execution state machine • UNCERTAIN state for ambiguous external outcomes • Deterministic policy enforcement • Append-only audit trail • Failure handling and reconciliation EVALUATION On a frozen 100-item benchmark evaluated across 20 random outcome seeds: RECOVER-ALLOC: ₹1,74,817.99 mean realized recovery Oracle: ₹1,92,135.66 mean realized recovery Realized Recovery Ratio vs Oracle: 90.99% RECOVER-ALLOC improvement: +25.89% vs Random +90.07% vs Static Rules +120.00% vs Blind Retry MODEL CALIBRATION I also evaluated selection-induced optimism caused by optimizing predicted probabilities. Raw model selected-portfolio bias: +9.80% After isotonic calibration: +0.68 percentage points This helped ensure that probability estimates remained more reliable after the optimizer selected the recovery portfolio. PROJECT GitHub: https://github.com/red-coder-27/recover-alloc Live Demo: https://recover-alloc.onrender.com/ This project was built as a solo submission for the Razorpay AI Buildathon 2026, AI Revenue Recovery track. #Razorpay #RazorpayAI #AIBuildathon #AI #RevenueRecovery #MachineLearning #Optimization #CP-SAT #MCMKP #Fintech #Python