Sentiment Analysis at Scale

Fine-tuned a BERT-base model on a dataset of 1M+ e-commerce product reviews. Implemented a custom training loop with mixed-precision and gradient checkpointing to fit within a single A100 GPU. Achieved 94% accuracy on the held-out test set, outperforming the DistilBERT baseline by 3 points. Deployed as a FastAPI microservice with async batching, serving ~500 req/s on a single instance.