From Sketch to Scale.

Developed advanced wildlife monitoring capabilities by fine-tuning YOLO and custom Transformer based architectures along with (CUDA, TensorRT) inference optimization. Built a robust backend infrastructure (FastAPI, PostgreSQL, Docker) featuring optimized queries and an end-to-end MLOps pipeline for continual learning from camera trap data. Mitigated annotation bottlenecks using Active Learning algorithms and ensured system health via custom API monitoring tools.

Enhanced zero shot inference capability of LLMs through supervised fine-tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). Curated and refined domain-specific datasets for complex reasoning tasks. Optimized reward models using human preference data to better align outputs with user expectations (truthfulness, harmlessness, instruction-following). Collaborated with ML engineers on refining annotation guidelines and feedback mechanisms.

Improved latency and throughput for live media transfer from semi-autonomous vehicles to edge servers under Dr. Arani Bhattacharya, utilizing state-of-the-art multipath QUIC protocols. Critically analyzed and tested Alibaba's XQUIC and Tencent's TQUIC frameworks to identify solutions for latency bottlenecks. Reported detailed findings on performance and build library inconsistencies.