Projects

An agentic tool that profiles how an LLM uses GPU memory on real hardware and produces ranked, quantified optimization recommendations — cutting inference cost without an ML-performance engineer on staff.

A multi-agent system that takes a CUDA kernel, profiles it with NVIDIA Nsight Compute on a real GPU, diagnoses the bottleneck (compute-bound vs memory-bound vs L1/load-throughput-limited), and uses an LLM to generate a faster rewrite — then proves the change is correct and measurably faster before accepting it. First verified run: a naive 1024×1024 matmul (1.185 ms baseline) was rewritten by the optimizer as a shared-memory-tiled kernel (0.936 ms, identical output) — accepted in a single iteration.

A public, stateless agentic AI article generator. Type any AI/ML topic; a LangGraph state machine (Research → Drafter → Critic → Verify with a retry edge) runs against an arxiv research-paper corpus and produces a sourced, grounded article — streamed live to the visitor and downloadable as PDF. Every agent has real tool use: the Researcher picks its own searches, the Drafter calls verify_claim mid-write to ground specific assertions before they ship, and the Critic produces a structured grounded/ungrounded verdict via forced tool output. Per-run cost capped, per-IP rate-limited, nothing saved.