ReconGrid AI
An open-source settlement reconciliation and discrepancy diagnostic engine with an audited Q&A agent for Razorpay merchants and finance teams.
Originally built for the Razorpay Buildathon 2026 (Track 04: AI Finance Controller), ReconGrid AI automates the painful process of matching bank statements against Razorpay payment gateway settlements.
Reconciliation in 60 Seconds (Why This Exists)
If you have never worked in fintech or accounting, here is the problem in plain terms:
Suppose a customer buys a ₹10,000.00 product on your website through Razorpay. You might expect your bank account to show a credit of ₹10,000.00. It never does.
Instead, your bank statement shows a credit of ₹9,664.00:
Gross customer payment: ₹ 10,000.00
- Gateway fee (2.0% MDR): - ₹ 200.00
- 18% GST on that fee (18% of ₹200): - ₹ 36.00
- E-commerce TDS (Section 194-O, 1%): - ₹ 100.00
────────────────────────────────────────────────────────
Net amount credited to your bank: ₹ 9,664.00
Now multiply that by 5,000 transactions a month. Add in:
- Batched payouts: Razorpay combines 40 customer payments into one single lump-sum bank deposit.
- Mid-cycle refunds: A ₹1,500 customer refund is deducted from tomorrow's settlement batch.
- Messy bank exports: HDFC, SBI, ICICI, and Axis export CSVs and PDFs with non-standard columns, encrypted passwords, and UTRs (Unique Transaction References) buried inside 80-character narration text.
Chartered Accountants (CAs) and finance leads spend 20 to 40 hours every month manually cross-referencing Excel sheets with VLOOKUP to find out where missing rupees went.
The ReconGrid Solution: Deterministic Math + Guardrailed AI
ReconGrid AI enforces a strict architectural boundary:
- 100% Deterministic Code for Money: All matching, fee deductions, tax calculations, and ledger balances run in pure Python using
Decimal (zero floating-point math). AI never touches, calculates, or decides financial amounts.
- AI Only for Plain-English Explanation: An LLM is used strictly to answer user questions about pre-computed audit facts (e.g., "Why did order #4521 have a ₹336 delta?").
- Regex Numeric Guardrail: An anti-hallucination check compares every number in the LLM's answer against the database record. If the LLM invents or alters any number, its text is immediately discarded and replaced with raw database facts.
Key Terms (Fintech Primer)
If you are new to Indian banking and payment gateways, here are the terms used throughout this codebase and documentation:
| Term | What It Means | Why It Matters Here |
|---|
| UTR | Unique Transaction Reference | The 12-to-22 character alphanumeric ID assigned by the banking system (NEFT / RTGS / IMPS) to identify a money transfer. This is our primary key for matching bank rows to gateway settlements. |
| MDR | Merchant Discount Rate | The fee Razorpay charges per transaction (typically ~2%). |
| GST on MDR | 18% Goods and Services Tax | Indian tax applied strictly to the gateway fee (e.g., 18% of ₹200 fee = ₹36 GST). |
| Section 194-O TDS | Tax Deducted at Source | Under Indian income tax law, e-commerce operators must withhold 1% tax on gross sales. |
| Paise vs. Rupee | $1\text{ INR} = 100\text{ paise}$ | Razorpay APIs handle money in integer paise (₹100.00 = 10000 paise) to avoid decimal errors. Bank statements report in rupees. Ingestion normalizes everything to Python Decimal(18, 2) rupees. |
| Conservation | Mathematical balance guarantee | $\sum(\text{matched} + \text{suggested} + \text{conflicts} + \text{exceptions} + \text{pending}) = \text{total ingested}$. Zero dropped rows, zero ghost rupees. |
Razorpay Hackathon Track 04: AI Finance Controller
ReconGrid AI was built specifically to address the mandate of Track 04:
"Build finance-ops agents that close the loop over synthetic data with 50+ record batches, reporting match rates and unresolved anomalies... Verification capacity, not generation speed, is the bottleneck."
This is the beginning of the README. Read the rest at the original. Read the full README ↗
