iivek/ledgerlens is an offline mobile project that extracts structured per-item data from receipt photos. The repository presents Python-based model work, a React Native app, a llama.cpp execution path, and evaluation/rescoring scripts. What the README introduces is a setup that handles a Qwen2-VL-2B-based model in a way suited to on-device processing.
The project links evaluation to distillation, which transfers the behavior of a large model to a smaller one, quantization, which reduces model size, and pruning, which removes some connections. It compares size and accuracy changes under compression for models that handle images and text together (VLMs), and provides SROIE and CORD measurement results along with tools to re-evaluate them. This setup is an attempt to look at model compression and on-device execution together in receipt extraction.
That said, the published SROIE and CORD measurements are a single run obtained on a specific Mac. The README also notes that the quantization comparison should not be treated as a definitive conclusion and that repeated measurements are needed. Mobile phone execution and additional compression evaluations are still in progress. Therefore the current results are closer to an evaluation stage that must be validated through repeated measurement and actual execution on a mobile phone, rather than a settled ranking of compression methods.