Built and pretrained a PyTorch MoE SLM on Gated DeltaNet 2, with multi-token prediction, fill-in-the-middle, and context-length curriculum. Includes cached Parquet loading, online sequence packing, custom Triton kernels, Muon + 8-bit AdamW optimization, and WSD scheduling. Trained on 1B+ tokens (130k vocabulary) to a pretraining perplexity of ~20 with closely matched train/val curves.
Conditional diffusion training framework with DDPM, mixed precision, cached text embeddings, DDIM sampling, and a custom U-Net. A second research pass added dynamic aspect-ratio batching, automated captioning, and resumable Jupyter-based training workflows.
Production RAG system for internal knowledge sources using FastAPI, LangChain, and Pinecone. Implemented document ingestion, vector embeddings, semantic retrieval, prompt orchestration, and REST deployment for structured and unstructured data.
End-to-end RTSP/IP-camera analytics: custom YOLOv4 training on AWS SageMaker, automatic license-plate recognition, multi-object tracking, and entry/exit monitoring. Achieved 45 FPS on GPU with 90% recognition accuracy after quantization.
Real-time IP-camera face recognition using rotated face detection, embedding extraction, and approximate-nearest-neighbour search with FAISS and Annoy for robust identity matching across viewing angles, on both CPU and CUDA.
Compression pipeline built on pretrained VQGAN / KL autoencoders with custom serialization and reconstruction, producing file sizes comparable to 20–30% JPEG compression at higher perceptual quality.
Android drowsiness and head-pose detection system using quantized TensorFlow Lite models with Java/C++ integration and OpenCV DNN for low-latency, on-device inference.
From-scratch PyTorch re-implementations of landmark papers spanning GANs, diffusion, segmentation, recognition, and knowledge distillation, built to deepen architecture-level understanding and reproducibility.
Interactive tool for visualizing how different tokenizers split and encode text, useful for inspecting vocabulary behavior across language models.
Visualization interface for retrieval-augmented generation chat pipelines, showing retrieved context, embeddings, and response grounding.
Node-based visual editor for building and chaining image processing operations interactively, without writing code.
Computer Vision
Detection, tracking, segmentation, pose & keypoint, depth estimation, image captioning, real-time RTSP pipelines
Language & Retrieval
Custom SLM pretraining & SFT, LangChain, Hugging Face Transformers, vector databases, hybrid search, reranking
Generative AI
Diffusion models, image editing, latent control, from-scratch and fine-tuning training pipelines
Deployment & Inference
Ollama, llama.cpp, ONNX Runtime, quantization, GPU inference optimization, FastAPI
Multimodal & Speech
Vision-language models, VQA, image-text joint embedding, speech-to-text / text-to-speech pipelines
Engineering
Python, C++, Java, PyTorch, TensorFlow, OpenCV, AWS (S3, EC2, SageMaker), Pinecone, Linux
B.Sc. in Computer Science and Engineering