Personal · R&D
AI Receipt OCR
Receipt OCR & document tools
A set of OCR + LLM tools for extracting structured data from Turkish receipts and documents.
ÖRNEK MARKET A.Ş.
Çankaya / ANKARA
04.11.202618:42
SÜT 1L32,50
EKMEK15,00
KAHVE 250G289,00
ELMA 1KG76,00
TOPLAM412,50
KDV37,50
receipt.jsonQwen2-VL · LoRA
{
"firm":"Örnek Market A.Ş."0.98
"date":"2026-11-04"0.99
"total":412.500.97
"category":"market"0.91
"currency":"TRY"0.99
}
- Role
- R&D · solo
- Period
- R&D
- Context
- Personal
- Status
- Experimental
Highlights
- 01
Built a multi-engine Turkish receipt OCR benchmark in Python (PaddleOCR, EasyOCR, Tesseract, TrOCR) that auto-selects the best result by confidence score.
- 02
Developed receipt-analysis web services (Flask) combining OpenCV preprocessing with local Ollama vision models / OpenAI GPT to extract firm, date, total, category and currency with confidence scoring.
- 03
Built a Node.js/Express OCR API (Tesseract.js + Ollama) and Expo document-scanner apps (BelgeTarayıcı / Docscan) with native edge detection.
- 04
Built an end-to-end vision-LLM fine-tuning pipeline (Apple Silicon / MLX) targeting a LoRA fine-tune of Qwen2-VL-2B for Turkish receipt → structured JSON: automated labeling, an MLX-VLM dataset (~2,800 receipts) and a field-level accuracy evaluation harness.
Architecture
- 01
Capture
Expo scanner with native edge detection
- 02
OpenCV
Image preprocessing
- 03
OCR ensemble
PaddleOCR, EasyOCR, Tesseract, TrOCR — picked by confidence
- 04
Vision LLM
Field extraction with Ollama vision / GPT
- 05
JSON
Firm, date, total, category, currency + confidence
- 06
LoRA · MLX
Qwen2-VL-2B fine-tune on ~2,800 receipts
Other projects
All projects →
PiscatoriumHobby project · 2026 – present
Piscatorium
Sea & freshwater angling platform (web & mobile)
A fishing map that started in Türkiye and now spans 34 countries: species, regions, fishing grounds, techniques, baits/lures and tackle recipes, with a condition-based bite-score engine, interactive maps and a catch-log community. Weather, spots and legal size/season rules on one screen.
- Laravel 13
- PHP 8.3
- PostgreSQL / PostGIS
- Redis
- Vue 3
Dare a PlanPersonal · published on iOS & Android
Dareaplan
Personal & family finance planner
Multi-currency budgeting app: income/expense tracking, savings goals, investments, shared family plans and AI receipt scanning.
- React Native (Expo)
- TypeScript
- Laravel 12
- MySQL
- Firebase