Deep Learning Research Engineer building and evaluating AI systems from scratch in PyTorch, with a focus on efficient language models, generative vision, and research tooling.
I work at the boundary of model architecture and systems constraints: memory-aware attention, mixture-of-experts training, state-space models, diffusion pipelines, and reproducible experiment infrastructure.
- HyMo — hybrid language-model research combining recurrent and attention-based components.
- FusionLLM — hybrid LLM pre-training framework integrating latent attention, gated delta networks, mixture-of-experts, and multi-token prediction.
- GPT-OSS-Lite — long-context mixture-of-experts language-model experiments.
- StableDiffusion — from-scratch diffusion-model training work.
- EE1101 — offline-first interactive textbook for Electronic Systems and Circuits.
Language-model systems — HyMo, FusionLLM, GPT-OSS-Lite, Mamba-3-Lite, LLaMA-3-Lite, DeepSeek-v3-Lite, TranslationLM, and GPT-From-Scratch.
Vision systems — StableDiffusion, detect-objects, upscale-sr, ActionRecognition, FaceAgingCycleGAN, FaceGenerationVAE, DCGAN-Face-Generation, and VisionLanguageModel.
Research and agentic tooling — AutonomousResearcher and news-agent.
Education systems — EE1101, an interactive course-textbook delivery system with an offline static build.
I favor inspectable implementations, numerical checks, repeatable experiments, and clear technical documentation. Each repository is the primary record for its architecture, validation method, and limitations.