AVAminVost
FA
← Back to projects
AIWebInfrastructure

Abzar Market OCR ai.abzarmarket.net

Production Persian OCR and price-list extraction platform for supplier documents, built around local vision-language/OCR models and seller-specific adapters.

What I did

Designed and implemented end-to-end, including dataset preparation, model fine-tuning, application architecture and production deployment.

Highlights

  • Built an RTL Next.js panel and Python/FastAPI processing backend with PaddleOCR PP-OCRv5 detection, real bounding boxes, assisted region selection, automatic row grouping and crop preview.
  • Created a custom dataset of about 1,167 samples and performed QLoRA fine-tuning for Qwen-based OCR/vision-language workflows.
  • Designed a base-model plus per-seller/per-brand adapter architecture, with separate text, price and SKU adapters managed through MySQL metadata.
  • Current internal evaluation reached more than 87% accuracy for price extraction and about 98% for SKU/product-name fields in the evaluated datasets.
  • Configured local GPU execution based on available hardware and delivered an initial production version.

Technology

Next.jsReactPythonFastAPIPaddleOCRQwen/Qwen-VLQLoRAMySQLTailwind CSSLocal AI ModelsGPU runtime