What is this project?

This project implements a production-ready NLP sentiment analysis pipeline. A domain-fine-tuned BERT-base model classifies customer reviews as positive, negative, or neutral with 94% accuracy. The async FastAPI backend handles both single-text and batch inference, while Redis queuing smooths out traffic spikes. The React dashboard visualises sentiment trends over time with interactive filters by product, date range, and confidence score.

The biggest engineering challenge was latency — naive BERT inference was 350ms per request. By switching to ONNX Runtime with 8-bit quantisation and implementing intelligent request batching, average latency dropped to under 80ms while GPU memory usage fell by 40%. The result is a system that can handle bursts of 500+ concurrent users without degradation.