I build production-grade AI across LLMs, computer vision, and generative AI, from training large-scale models to shipping optimized inference.
Experience
Research and development of generative AI models for 3D. Scaled distributed training across hundreds of GPUs with DeepSpeed ZeRO and optimized inference for cost-efficient deployment, reducing GPU memory and latency. Deployed production 3D generative AI models on Ray across multimodal inputs (text, image, sketch, point cloud).
Researched Explainable AI methods (Integrated Gradients, SHAP) for DNA sequence and GNN healthcare models using PyTorch Geometric on HPC clusters. Improved accuracy by 20% through hyperparameter tuning and cross-validation.
Built real-time video analytics (detection, tracking, re-ID, OCR, action recognition, pose estimation) in Python/C++, driving 4 new smart city contracts. Optimized multi-GPU pipelines with TensorRT and DeepStream, achieving 2x faster training with distributed fp16 mixed precision and 3x inference throughput via INT8 quantization.
Computer Vision R&D Intern · Jun 2019 – Jul 2020
Multi-label Pedestrian Attribute Recognition via image processing and clustering. Monocular depth estimation for under-vehicle detection with RGB-D data. Semantic/instance segmentation for Indian road scenes (U-Net, DeepLab, Mask R-CNN).
Applied NLP techniques (NLTK, scikit-learn) on resumes for automated information extraction, classification, and entity recognition.
Education
Graduate Teaching Assistant for Data Mining, Data Science, and Programming with Python.