EgoScan: Paper to Ledger

EgoScan is an open-source AI bookkeeping engine that turns photographs of handwritten shop notebooks into structured sales records, customer debt lists and daily reports.

Many small traders in Nigeria still record transactions in exercise books. Manually transferring those records into digital bookkeeping software takes time, while generic OCR can struggle with handwriting, naira shorthand, payment marks and local bookkeeping conventions.

EgoScan uses Gemma 4’s vision capabilities through the Gemini API to transcribe notebook pages into structured records. Deterministic Python code then parses amounts, recomputes quantity × price, checks extracted totals against the transcribed lines and written page total, and flags doubtful fields for human review. Only confirmed pages contribute to the application’s debts, reports and exports.

What makes it different

• AI with financial checks: The model transcribes; application code performs calculations and flags inconsistencies.
• Built for shop notebooks: Handles Nigerian naira shorthand, customer headings, payment marks, dated entries and struck-through debts.
• Human-reviewed records: Traders can inspect the original photograph, correct extraction errors and confirm pages.
• Open-source infrastructure: Includes a MIT-licensed engine, a versioned extraction API, CSV and Excel exports, and integration with EgoTrack’s merchant data model.

Technical validation

In a benchmark using 12 synthetic notebook pages, EgoScan found 99.25% of expected rows. With image strips, 94% of matched rows had exact totals, and the validation checks flagged all wrong totals in that test.

A separate development run processed 22 photographs of real shop notebooks, yielding 209 rows. Independent ground-truth evaluation of real-handwriting accuracy remains outstanding.

Technology

Python, FastAPI, SQLite, Gemma 4, Gemini API, Pillow, HTTPX and openpyxl.

Repository: https://github.com/nevodesigns/egoscan

Security and engineering

I attacked the app with a 34-check penetration test and fixed what it found (10 failing checks before, 0 after). Static analysis found 0 issues, a vulnerability lookup over 26 dependencies found none, and the git history contains no secrets. A written assessment lists what is still open. 106 automated tests, an 18-step browser smoke test and CI run on every commit. Verified against EgoTrack’s real merchant store: scanned records import and the store’s own totals match.