The Great IndianFDE Hackathon 2026

The fastest route in 2026. Open Claude Code, Cursor, or any agentic coding tool in a folder containing the downloaded data, and paste this brief.

You are competing in The Great Indian FDE Hackathon. This folder contains
vendor_invoices.csv, purchase_orders.csv, vendors.csv, products.csv,
sample_submission.csv, and four policy PDFs from a fictional manufacturer
called Deshmukh & Rao Industries.

Build a pipeline that classifies every invoice in vendor_invoices.csv into
exactly one of: CLEAN, QTY_MISMATCH, RATE_MISMATCH, DUPLICATE_INVOICE,
MISSING_PO, VENDOR_MISMATCH, GST_ERROR, UOM_MISMATCH.

Join invoices to purchase orders on po_number. Use deterministic checks
where arithmetic decides: quantity vs PO, rate vs PO, GST at 18%, whether
the PO reference exists, duplicate billing against one PO (the later-dated
invoice is the duplicate), unit-of-measure differences such as Box(10) vs
Nos, and the vendor code on the invoice vs the PO. Where records do not
line up cleanly — vendor names in different formats, ambiguous references —
use model calls or fuzzy matching, and comment your reasoning.

Produce submission.csv with columns invoice_number,status, covering every
invoice exactly once, matching sample_submission.csv's shape. Then report
the class distribution you predicted and the cases you were least sure
about. The metric is macro-F1 across the eight classes, so the rare classes
matter as much as CLEAN.