An unmanned-retail vision-AI system that recognizes products by camera and carries them from price lookup through checkout. Book recognition is implemented first, with a design that extends to packaged and refrigerated goods.
Built on SigLIP2 embeddings, CSLS retrieval, FAISS, and EasyOCR, it targets open-set recognition where new products are added by enrollment alone — no retraining. On the demo catalog it reaches 98.6% Recall@1 and 90.3% safe auto-checkout at 0% mis-charge / false-accept, with ~150ms p95 latency.
Final project of a generative-AI multimodal vision course · solo build · Jul 2026 (~4 weeks). The full deck is embedded below.