The Problem
Living in Japan means getting handed receipts everywhere — convenience stores, pharmacies, restaurants. Most end up in a pocket or trash, never tracked, and the coupons go unused.
The Solution
Sysl is a PWA that scans any Japanese receipt automatically. Point the camera, tap once, and the store name, items, total, payment method, and coupons are extracted and saved — no manual entry needed. It installs directly on your phone home screen.
Every receipt gets pinned on a community map so users can see what people nearby are buying and what coupons are available at local stores.
Databricks
Every scan logs two MLflow runs to the receipts-ocr experiment on Databricks Free Edition — one for OCR quality metrics (confidence score, block count, low-confidence blocks), one for extraction results (store name, category, total spend, payment method, coupons found). Across all scans, over 30 runs are logged and tracked.
[Image 1 — MLflow runs list]
After 17 real receipts scanned across convenience stores, groceries, restaurants, and bakeries in Japan:
- Total spend tracked: ¥7,786
- Average per receipt: ¥458
- Average OCR confidence: 87.16%, peak 99.81%
[Image 2 — summary table]
The analytics notebook breaks down spending by category, receipt count by store type, and OCR confidence trends across scans.
[Image 3 — spend by category charts]
Stack
Next.js, FastAPI, PaddleOCR PP-OCRv5, GPT-4o-mini, Supabase, Mapbox, Databricks Free Edition
What's Next
Predictive budgeting on Delta Lake, coupon recommendations via collaborative filtering, and Databricks AI Gateway for model governance and A/B testing.
Video: https://youtu.be/Cg6nmgnvwGY
Uploaded by beepBop on 2026-06-05.
