Results

A hand-written Triton kernel takes a memory-bound op from PyTorch's 11% bandwidth utilization to 88%, an 8.1x speedup.

A pure-Triton Mixture-of-Experts kernel matches or beats Stanford's CUDA-optimized Megablocks at inference batch sizes, on both NVIDIA A100 and AMD MI300X, with 35% less global memory traffic and no CUDA.

Tree speculation, EAGLE, and Medusa combined, with reference implementations and an acceptance-rate model, on memory-bound autoregressive decode.

Linear probes on hidden states detect deliberate underperformance across Mistral, Gemma, and Qwen, before any output is generated. Steering cut Gemma sandbagging by 20%.

Distributed, statistically rigorous LLM evaluation: bootstrap confidence intervals, paired significance tests, and content-addressable caching on Delta Lake.

Quality-diversity (MAP-Elites) red-teaming surfaces diverse, transferable vulnerabilities. A mid-generation model can be more vulnerable than both its predecessor and its successor, a pattern static benchmarks miss.

On a 34-task TMDB corpus, a semantically planned tool surface and a one-tool-per-operation baseline both scored 28/34, while composite tools cut total calls from 312 to 289 at the cost of about 20% more context bytes. Four pre-registered task-success comparisons were underpowered to resolve a difference.