Work
Selected work in post-training, agents, inference systems, and computer vision.
Post-training
Mathematical Reasoning
A Qwen2.5-3B SFT and Dr. GRPO pipeline evaluated on GSM8K and MATH500.
Coding agent
Sandboxed Coding Agent
A LangGraph CLI with repository-scoped E2B sandboxes and LangSmith traces.
Inference runtime
Inference Engine from Scratch
A compact single-GPU engine with paged KV cache, block allocation, continuous batching, and H100 measurements.
Computer vision
Diffusion Image Super-Resolution
A 4× DIV2K super-resolution model combining a time-conditioned U-Net, Swin attention, and residual diffusion.