Autonomous Transaction Reconciliation at a Pan-African Payment Processor

The challenge

The client — a fast-growing payment processor operating across 12 African markets — was spending over ₹1.74 Cr/month on a manual reconciliation team. The team of 14 struggled to hit end-of-day deadlines during peak volumes, resulting in settlement delays, regulatory escalations, and significant operational risk. Off-the-shelf reconciliation software couldn't handle the complexity of multi-currency, multi-rail transactions.

Our approach

1. 1

   Ran a 2-week discovery to map all transaction flows, exception types, and escalation rules.

2. 2

   Designed a 3-agent LangGraph pipeline: an Ingestion Agent (normalises data from 6 sources), a Matching Agent (rule-based + ML hybrid), and an Exception Agent (classifies, resolves, or escalates unmatched transactions).

3. 3

   Deployed on Kubernetes with Airflow orchestration; dbt used for all downstream transformations.

4. 4

   Built a real-time dashboard for ops staff to monitor exception queues and approve escalations.

5. 5

   Ran 4-week parallel run alongside the human team before full cutover.