Production RAG with FastAPI, LangChain, Pinecone, and OpenAI. Hybrid retrieval for structured + unstructured data, prompt orchestration, conversation memory, and AWS deployment. Built for scale with observability.
End-to-end video analytics: YOLOv4 detection, ALPR, OCR, DeepSORT tracking, and re-identification. 40-45 FPS on GPU, 90% recognition accuracy, 4× model compression via quantization/distillation.
Custom DDPM implementation in PyTorch with mixed precision, gradient accumulation, cosine schedule, and DDIM sampling. Trained on custom datasets with thorough logging and evaluation.
VQGAN-based compression pipeline with learned codebook, latent storage, and 8-bit serialization for efficient archival. Perceptual loss + adversarial training for high fidelity.
Clean, from-scratch PyTorch re-implementations of landmark papers: GANs, diffusion, segmentation, and knowledge distillation. Focus on architectural understanding and reproducibility.
Computer Vision
Detection, Tracking, Segmentation, Pose, Depth, OCR/ALPR, Video Analytics, 3DGS
AI / ML
PyTorch, TensorFlow, HuggingFace, RAG, LLMs, Diffusion, GANs, Transformers, VLM
Engineering
Python, C++, Docker, FastAPI, ONNX/TensorRT, AWS/GCP, Linux, MLOps