← All projects

Project 05 Planned

GSTR-2B vs Purchase Register Reconciler

Every accountant reconciles GSTR-2B against the purchase register each month. Mismatches mean lost input tax credit.


For: Indian SME CFOs and CA firms · Planned — Month 5 (flagship).

The problem


Every Indian accountant reconciles GSTR-2B against the purchase register each month. It is high-volume, deadline-bound, and a missed mismatch means lost input tax credit — real money.

What it does


Upload the GSTR-2B JSON and the purchase register Excel. The tool returns matched, value-mismatched, GSTIN-mismatched, missing-in-books and missing-in-2B buckets, each with a confidence score. For the mismatches, an LLM drafts the vendor follow-up email.

  • Fuzzy vendor name matching across legal-entity variants
  • GSTIN checksum validation and date tolerance window
  • Downloadable reconciliation Excel

Design choices


The matching has to survive real-world mess: the same vendor as Reliance Retail, Reliance Retail Ltd and RRL; dates a few days apart; GSTINs that need checksum validation. The deterministic matching does the reconciliation; the LLM is used only for the email drafting, where a wrong word costs nothing and a wrong number would cost ITC.

The synthetic GSTR-2B is built so all five reconciliation outcomes appear — a reconciler tested only on clean data is a toy.

Stack


PythonpandasClaude Haiku

Built and demonstrated on synthetic data only.