HFT-X is a high-frequency trading engine featuring a custom UDP market data feed capable of streaming 100K+ msgs/sec into a real-time order book and matching engine. The system achieves sub-1ms tick-to-trade latency (p99) using optimized in-memory data structures, non-blocking I/O, and microsecond-level event processing.
Built a multi-head YOLOv8 model for joint obstacle detection and pavement segmentation, cutting latency by 33%. FastDepth on Jetson Nano for real-time monocular depth-8FPS, context-aware audio navigation - visually impaired users. Achieved 85.2% mAP (obstacles) and 85.9% mAP (curb segmentation) | Patent application in progress.
Designed a modular, async drift detection framework unifying statistical and ML-based detectors, reducing false drift alarms by 38% and improving true drift detection by 27% over single-method baselines. Optimized feature-level explainability by selectively executing SHAP only post-drift, cutting explanation latency by 61%