Case study·AI invoice processor

ZeroEntry — 40 seconds per invoice, end-to-end.

Vision LLM + OCR fallback + classification + Tally posting + GSTR-2B reconciliation. Built for Indian CA firms and AP teams drowning in manual invoice entry.

Live · production
ZeroEntry product UI screenshot
40s
per invoice
vs 7-min manual entry
90%+
extraction accuracy
machine-printed
10×
cycle speedup
vs hand-keying
The problem

What we set out to solve.

Mid-size Indian CA firms process 200–800 invoices/day across clients. At 7 minutes each manually, that's 23–93 associate hours daily. Errors compound at GSTR-2B reconciliation. Hiring more associates doesn't scale — unit economics break past 50 clients.

The build

What we actually shipped.

01 of 05

Vision LLM extraction

Gemini 2.0 Flash / Claude 3.5 Sonnet Vision for primary extraction. Pulls vendor name, GSTIN, line items, tax breakdown, totals, dates.

02 of 05

OCR fallback

Tesseract for stamped, scanned, handwritten invoices the Vision LLM struggles with. Confidence-routed.

03 of 05

Classification engine

B2B / B2C / RCM / export-LUT / capital goods auto-classified. ITC eligibility flagged.

04 of 05

Tally XML posting

Posts to Tally Prime via XML connector. Audit trail per entry. Vendor master deduplicated against PAN/GSTIN.

05 of 05

GSTR-2B reconciliation

Auto-fetches 2B per client, matches against posted entries, flags mismatches with specific reason codes (rate diff, ITC ineligible, missing in 2B).

The stack

Built on the same Indian-stack-native pieces we ship every day.

Gemini VisionClaude 3.5 Sonnet VisionTesseract OCRTally XMLZoho Books APIPostgresS3 (encrypted)
The outcome

How it's playing out.

Deployed across CA firms and internal AP teams. Replaces 60–80% of manual data entry. GSTR-2B mismatch resolution moved from Saturday-night to a 30-min Monday morning review. Foundation for the productized GSTR-2B reconciliation SKU.

We went from 14 hours a week of invoice entry to under 2. The mismatch flag report is what I want from every tool — specific, actionable, not generic.
Reference customer — Mumbai CA firm

Want this playbook for your business?