Information Extraction from Receipts

At Blackstraw I worked with the client's team on extracting every item - description, quantity and price - from photos and scans of printed receipts, despite heavy perspective distortion, creases and varied layouts.

The pipeline chains several computer-vision models (text detection, OCR, line detection) with NLP models that structure the result: text classification, named-entity recognition, coreference resolution, dependency parsing and knowledge-graph embeddings for product matching.

A key component is a lightweight Graph Attention Network that groups OCR detections into text lines, which became a published paper. Models were served with TF Serving and TensorRT, with custom CUDA kernels via PyCUDA.