GPU systems engineer. I care about the layer where hardware meets model: correctness, reliability, and the infrastructure that makes AI compute trustworthy.
CUDA kernels don't throw exceptions when they're wrong. This harness surfaces silent failure modes: warp-shuffle reductions returning half the sum, softmax overflowing to NaN on large logits, tile boundaries that break at N=257. Kernels compiled via nvcc, loaded into Python through ctypes, validated against NumPy CPU baselines.
Built and validated a Flash Attention Triton kernel from scratch. Found a non-obvious failure: padding slots in the last K-tile load as 0.0 instead of -inf. exp(0)=1 sits in the softmax alongside real scores. At N=1, that's 63 fake ones against one real score output comes out 64x too small. No crash, no warning, just wrong. Built a harness around that and three other failure modes: tile boundary edge cases, causal mask integrity across block boundaries, and FP16 drift at long sequences.
On dissolving the noise of the mind and finding stillness in the present moment.
A lone astronaut, an impossible problem, and the most unlikely friendship in the universe.
On why people think and communicate so differently, and how to actually reach them.