Reconly

Agentic AI Payment Reconciliation Platform autonomously reconciles high-volume payments using deterministic rules and generative AI us AWS NOVA.

Comment

Inspiration

Enterprise payment reconciliation remains heavily manual, rule-fragile, and slow, especially across modern real-time rails like UPI, RTP, cards, and open banking. We were inspired by the gap between traditional deterministic matching engines and the growing potential of agentic generative AI to reason, explain, and adapt. Our goal was to demonstrate how next-generation AI agents-working alongside proven rules-can transform reconciliation from a reactive back-office task into an intelligent, self-optimizing system.

What it does

Agentic AI Payment Reconciliation Platform autonomously reconciles bank statements against ledger invoices using a multi-agent pipeline:

  • Ingestion Agent loads statements and ledger data
  • Deterministic Matching Agent performs fast, exact matches
  • AI Matching Agent reasons over ambiguous cases
  • Exception Agent classifies and recommends actions

The platform produces explainable matches, confidence scores, and actionable exceptions through an interactive dashboard.

How we built it

  • Backend: Python + FastAPI microservices
  • Agents: Deterministic matcher, LLM matcher, exception analyzer
  • Models: Amazon Nova Lite (via AWS Bedrock) and local fallback mode
  • Frontend: React dashboard with pipelines, KPIs, charts, and tooltips

We implemented an agent orchestration flow that mirrors real-world financial operations and supports both local simulation and cloud-based inference.

Challenges we ran into

  • Ensuring reliable JSON output from LLMs
  • Balancing deterministic precision with AI flexibility
  • Preventing unnecessary AI calls when rules already succeed
  • Designing UI feedback that clearly explains why a match occurred

Accomplishments that we're proud of

  • Hybrid deterministic + agentic AI architecture
  • Explainable AI reasoning embedded into UI tooltips
  • Local vs AWS execution toggle
  • End-to-end working reconciliation demo

What we learned

  • Agentic patterns outperform single-call LLM approaches
  • Explainability is as important as accuracy in financial AI
  • Small prompt design changes can drastically improve results

What's next for Agentic AI Payment Reconciliation Platform

  • Learning feedback loops to improve matching over time
  • Multi-ledger and multi-currency support
  • Integration with ERP and payment gateways
  • Human-in-the-loop review workflows
  • Enterprise-grade security and audit trails

Our vision is to evolve this into a fully autonomous financial operations copilot.

Built With

Share this project:

Updates