mrnirmata

AI systems engineer. agents, inference infra, research.

self-taught. 7 years shipping software.

i build AI agents and the harnesses that keep them honest. when the bottleneck sits lower in the stack, i go there. lately that meant GPU kernels.

work

  • lossless-bf16

    BF16 weights stored in 12 bits, bit-exact. 1.1–1.3x faster than cuBLAS on big layers at batch 1–16, H200.

  • h200-decode-notes

    what actually limits LLM decode on an H200. measured, including the ideas that didn't work.

what i'm good at

  • agent harnesses that are hard to game. verifier-gated loops, adversarial review, held-out evals
  • many agents on one long problem, without them stepping on each other
  • LLM inference: MoE, attention, speculative decoding, weight formats
  • GPU kernels down to SASS. TMA, wgmma, clusters, PDL. i count issue slots when it matters
  • tools that test themselves. bit-exact gates, automatic fallbacks, GPU work without a local GPU
  • measurement. most benchmarks lie a little. mine tell you how