Overview
For a distribution business that wanted every cheque in their Busy accounting software without staff learning it. An employee photographs a cheque on WhatsApp; the bot extracts amount/date/party/bank with GPT-4o Vision, matches the party (including learned alternative names), fetches unpaid bills, and writes the receipt voucher into Busy.
It was a team project; the WhatsApp cheque bot was mine to build and own.
Key Features
Cheque processing
- GPT-4o Vision OCR with orientation handling; documented 10+ OCR approaches to get handwriting right.
- Queue-based bulk processing — send a stack of cheques at once.
- Accuracy is measured from operator approve/disapprove feedback in the dashboard.
Quotation & catalog
- RAG catalog search with LangChain/LangGraph + pgvector + OpenAI embeddings.
- Busy integration over an RDP/SQL-Server bridge for parties, bills, and vouchers.
Tech Stack
- FastAPI
- OpenAI
- Evolution API
- PostgreSQL
- pgvector
- LangGraph
- Redis
- GCP
- Nginx
Architecture
- FastAPI + PostgreSQL/pgvector + Redis
- WhatsApp via Evolution API
- Deployed on GCP Compute Engine behind Nginx
Results
~95%+ overall; ~99% catalog/party
5–30 min
photo on WhatsApp
no accounting-software training
YourMitra
A live AI storybook platform where a child becomes the hero of their story — built end to end as the sole engineer, with paying customers.
← ExploreJain Parichay
Jain Parichay's matrimony platform. v1 (biodata-PDF circulation) is live; v2 — a full NestJS/Astro rewrite with an automated biodata-import pipeline — was in development. I was the software developer who built and ran it.
Explore →