README.md

metadata

title: Optimised Amazon RecSys
emoji: 🐢
colorFrom: pink
colorTo: gray
sdk: gradio
sdk_version: 6.4.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: HNSW Quantisation ONNX supported Optimised Amazon RecSys

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

• Engineered a Two-Tower (Bi-Encoder) semantic retrieval system (MPNet), fine-tuned on 826k+ user–item interactions using supervised contrastive learning (Multiple Negatives Ranking Loss), achieving Recall@10 of 0.2946 and strong query–product alignment.

• Scaled vector search from linear O(N) to logarithmic O(log N) by replacing brute-force FAISS IndexFlatIP with FAISS HNSW approximate nearest neighbor indexing, enabling real-time Top-K retrieval.

• Accelerated inference by 5× (32 ms → 6 ms) by migrating from PyTorch FP32 to ONNX Runtime with Int8 quantization, achieving a 4× reduction in model size and enabling cost-efficient CPU deployment.

• Optimized data and training pipelines using Automatic Mixed Precision (AMP) and implemented O(1) binary metadata caching, ensuring instant retrieval of multi-modal assets (images and videos) without data loading bottlenecks.

• Deployed a production-ready recommendation microservice on Hugging Face Spaces with a custom A/B benchmarking dashboard, validating the 5× latency reduction against the baseline model in real time.