Back to work
AI Document Automation Client Work In Production

Veridock

AI document automation built for Indian CA firms. A privacy-first pipeline that turns scanned invoices into validated, vendor-matched ledger data — 9 phases complete, 150 tests passing, running in production.

By the Numbers

The Problem

Chartered Accountancy firms drown in manual document entry. Every invoice that arrives gets read, keyed in, and cross-checked by hand — slow, error-prone, and impossible to scale during filing season.

The constraint that shaped everything: privacy. This is Indian client financial data, so the system was built privacy-first from day one rather than bolted on later.

Architecture

A FastAPI backend scaffolds the whole system: PostgreSQL via Supabase (7 tables), JWT authentication through fastapi-users, and Brevo webhooks feeding documents into the pipeline. Celery workers backed by Redis handle async processing, with uploads stored in Cloudflare R2.

The core is a 3-layer OCR pipeline: OpenCV preprocessing cleans the scan, Claude Haiku extracts structured fields, and a GSTIN validation layer verifies the output against India's GST identifier format. Extraction is never trusted blindly — it is validated before it lands in the ledger.

Key Features

Tech Stack

Python · FastAPI · PostgreSQL/Supabase · Celery · Redis · OpenCV · Claude Haiku · Cloudflare R2 · Brevo