Neurologist

gradient descent, from ring 0.

I work on the whole stack between the PCIe bus and the attention head.

$ cat /proc/neurologist
name        : Neurologist
role        : AI/ML engineer · systems programmer · VFX 
arch        : x86_64 + CUDA + whatever the problem requires
location    : Los Angeles, CA
languages   : Python, C++, Rust, TypeScript, CUDA C
philosophy  : if it's slow, profile it. if it's magic, disassemble it.
uptime      : shipping since the GTX 900 series

/proc/self/maps — where the work lives

ADDRESS RANGE        PERM   REGION
──────────────────────────────────────────────────────────────────────
0x0000_llm_engines   r-x    LLM inference — vLLM, llama.cpp, KV-cache
                            tricks, speculative decoding, quantization
0x0001_diffusion     r-x    SDXL / Flux pipelines, ComfyUI graph
                            surgery, LoRA training, VFX integration
0x0002_agents        rwx    multi-agent orchestration — AutoGen,
                            LangChain, tool-use protocols, evals
0x0003_dma_systems   r--    direct memory access, PCIe, memory-mapped
                            I/O, depth-buffer extraction, kbd/HID
0x0004_edge          r-x    models on hostile hardware — Pi 5, Jetson,
                            anything with a heatsink and a dream
[heap]               rw-    grows toward whatever's interesting
[stack]              rw-    C++ ↔ Python ↔ Rust, in that order

Selected work

Hush-Rivals-DMA	DMA-based hardware input pipeline — reading and injecting HID state over PCIe, zero software footprint on the target
dma-depth-buffer	GPU depth-buffer extraction via direct memory access — real-time scene reconstruction without touching the render pipeline
neural-inference-bench	Reproducible LLM inference benchmarking — tokens/sec, TTFT, and memory curves across vLLM, llama.cpp, and TensorRT-LLM
diffusion-pipeline-utils	Production utilities for Stable Diffusion / ComfyUI — latent-space tooling, batch schedulers, VRAM-aware pipeline planning
Licensing	Self-hosted license API + client SDK — HWID binding, offline grace periods, tamper detection

Toolchain

                       

Signal

opinions held strongly, benchmarked weakly-supervised.
the best abstraction is the one you can still read a hexdump through.

_{building things that think — and knowing exactly which registers they think in · LA}