Machine Learning Engineer | Computer Vision | Recommendation Systems | Edge AI

Architected a two-tower neural retrieval system in PyTorch with independent user and item embedding towers trained on implicit feedback signals, reducing candidate retrieval latency by 35%. Incorporated a Graph Neural Network layer over user-item interaction graphs improving recommendation recall@10 by 18%. Integrated an LLM-based semantic reranking stage raising click-through rate by 12%. Constructed feature pipelines across 5+ feature categories in Python and SQL cutting training data preparation time by 50%.

Orchestrated an async inference pipeline across N concurrent camera streams using YOLOv8-Pose detection, DeepSORT multi-object tracking, anomaly scoring, and WebSocket alert dispatch packaged as a thread-safe containerised FastAPI microservice. Optimised YOLOv8 via ONNX export and TensorRT FP16 quantisation achieving 30+ FPS on GPU edge hardware. Deployed via Docker Compose on AWS EC2 and on-premise edge nodes with a React dashboard featuring live feed grid, incident replay, and spatial anomaly heatmap.

Engineered a closed-loop automation pipeline where YOLOv8 severity scores trigger real-time Arduino C++ firmware responses including LED alerts, buzzer alarms, and motor actuation. Streamlined deployment across offline Edge Mode and Cloud Batch Mode. Established a PostgreSQL backend for detection event logging and trend analytics with ROI filtering to suppress false positives.