Builds

GPU kernels, inference engines, training pipelines, and agents. Each project ships benchmarks and code you can run.

2026 H100

Triton Kernels for LLM Inference

Triton · FlashAttention-2 · CUDA · PyTorch

Seven kernels from scratch — vector add through FlashAttention forward — each with a PyTorch baseline and H100 benchmarks.

16.9× FA speedup 257× less memory 85% of cuBLAS
2026 H100

mini-vllm

PagedAttention · Continuous batching · Qwen2.5-7B · CUDA

~500-line inference engine: paged KV cache, block allocator, and continuous batching. Same ideas as vLLM, one file per concept.

2.6× throughput 3.5× more seqs 160× max TTFT
2026

reasoning

Dr.GRPO · FSDP · PyTorch · Qwen2.5-3B · GSM8K

DeepSeek-R1-style post-training in raw PyTorch: SFT on OpenR1-Math CoT traces, then Dr.GRPO with exact-match rewards. No TRL, no HuggingFace Trainer.

GSM8K 78.2% Dr.GRPO 2× Blackwell FSDP
2026

coding-agent

LangGraph · vLLM · Qwen2.5-7B

Plan-then-execute CLI agent for real git repos. Self-hosted on vLLM — no API keys, no rate limits.

plan → tool loop self-hosted inference LangSmith tracing
2025

AI Agent Framework

MCP · FastAPI · HuggingFace Spaces

Multi-step tool-use agent built from scratch. Sliding-window memory, MCP tools, and evaluated on the GAIA benchmark.

GAIA benchmark MCP tools live demo