reconAI

Financial reconciliation can become difficult when businesses process thousands of payment, settlement, refund, and bank transactions. Even a small mismatch can require significant manual investigation. We wanted to build a system that could not only find these mismatches, but also help explain why they happened and what should be done next. This inspired razorRecon AI — an autonomous AI-powered reconciliation system for payment and settlement operations. razorRecon AI connects payment records, settlement records, and bank transactions and automatically checks whether they match. The system first uses deterministic financial logic to perform reconciliation. It identifies cases such as:

Missing transactions
Duplicate records
Amount mismatches
Delayed settlements
Incorrect transaction references
Partial settlements
Unexpected adjustments When an exception is detected, an AI investigation agent analyzes the available evidence and generates an explanation. Higher-risk cases can then be routed to a human reviewer instead of being automatically resolved. The system also maintains an audit trail so that financial teams can understand what happened and track the actions taken. We built the project using Python, with a Streamlit dashboard for the user interface and structured transaction data for the reconciliation workflow. The core workflow is:

Payment Data
     ↓
Settlement Data
     ↓
Bank Transactions
     ↓
Deterministic Reconciliation
     ↓
Matched / Exception
     ↓
AI Investigation Agent
     ↓
Risk Assessment
     ↓
Human Review
     ↓
Audit Trail

A key design decision was to keep financial calculations and reconciliation rules deterministic. The AI agent is used for investigation, explanation, and recommendations rather than inventing financial results. For the hackathon, we also used AO throughout the development process to coordinate AI coding agents and help build different parts of the application. One of the biggest things we learned was that building an AI application is not only about using an LLM. Reliable AI systems need clear boundaries between deterministic logic and AI reasoning. We learned how to:

Design an agent-based workflow
Connect AI agents with software tools
Handle structured financial data
Build exception-handling workflows
Keep AI responses grounded in available evidence
Add human review for high-risk decisions
Test AI-assisted workflows using measurable metrics
Use AO to coordinate AI-assisted development One of our main challenges was making the system reliable when transaction records were incomplete or inconsistent. For example, a payment might exist in one data source but not another, or the settlement amount might differ from the expected amount. We needed to make sure the system did not simply guess the reason for the mismatch. We addressed this by separating the reconciliation logic from the AI investigation layer. The reconciliation engine determines what is wrong, while the AI agent uses the available evidence to explain what might have happened. Another challenge was designing the agent workflow so that important or uncertain cases are sent to a human instead of being handled automatically. RazorRecon AI aims to reduce the manual effort involved in financial reconciliation and help finance teams investigate exceptions faster. Instead of simply showing > "Transaction mismatch detected." the system aims to provide: > "Mismatch detected → evidence identified → explanation generated → risk assessed → recommended action." Our goal is to demonstrate how autonomous AI agents can work together with deterministic financial systems to make reconciliation faster, more transparent, and more reliable.