~/work / hisabsaathi
HisabSaathi
Multimodal AI bookkeeping for Indian contractors and MSMEs. It turns WhatsApp UPI screenshots and messy bank-statement PDFs into clean, project-wise ledgers.
- document-understanding
- multimodal-extraction
- data-pipeline
01 / problem
Why it needed to exist
Contractors and small businesses run their money through WhatsApp payment screenshots, paper bahi-khata registers and multi-page bank PDFs. Every evening goes into matching them by hand, and money from different sites gets mixed up.
02 / what i built
The approach
01
Screenshot Se Hisab
A multimodal model reads any GPay, PhonePe or Paytm screenshot, printed bill or petrol slip and extracts the amount, UTR / reference ID, party name, date and payment direction. The entry is auto-assigned to a project and category.
02
Bank Statement Intelligence
Parses multi-page statement PDFs from SBI, HDFC, ICICI, Axis, PNB, Bank of Baroda, Union Bank, Canara and Kotak. It extracts every debit and credit, turns cryptic bank remarks into readable party names, and drops entries already in the ledger.
03
Offline-first by design
Construction sites have bad network, so OCR runs on-device and entries are cached locally, then synced to the cloud when connectivity returns.
04
Privacy as a feature
Uploaded slips are processed ephemerally for extraction only. Ledgers are encrypted, backed up and tied to an OTP-verified phone number.
03 / architecture
How data flows
- 00InputUPI screenshot · bill · bank PDF
- 01On-device OCRworks offline on site
- 02Multimodal extractionamount · UTR · party · date · direction
- 03Normalize + dedupclean remarks, match existing entries
- 04Categorizeproject + category assignment
- 05Ledger + reportsFirebase · WhatsApp PDF statements
04 / results
What it did in the real world
- Live on Google Play since October 2026
- Statement parsing for 9 major Indian banks
- Multi-project ledgers with live balances across banks, cash and wallets
05 / limitations
Honest notes
- Android only for now; iOS is next.