Writing
How I spend my Friday nights. Build logs, benchmarks, and lessons from real systems.
- Teaching Jev, a Decision Model, to Fly U.N. Squadron from Raw SNES MemoryUsing TypeSafe's Jev decision model to play a 1991 shooter: decoding the game's object table, freezing the emulator while the model thinks, and the request redesign that took median decision confidence from 0.27 to 0.85.Sep 2026
- Pooling a Tablet and a Workstation Over Thunderbolt to Run a 177B ModelWiring an RTX A6000 to a Ryzen AI Max+ 395 tablet over USB4 to run Qwen3.8-Flash-Next at full 262K context, and the control run that showed the cable made it slower.Sep 2026
- TensorRT-LLM on an RTX A6000: Part 2 — Concurrency, TTFT, and the Benchmark TrapPushing TensorRT-LLM under load on an RTX A6000: why synthetic prompt size misleads, why TTFT is a cleaner signal, and what it shows about Dynamo.Jun 2026
- TensorRT-LLM on an RTX A6000: Part 1 — GPU Passthrough, KV Cache, and Single-GPU ServingServing Qwen2.5-7B-Instruct with NVIDIA TensorRT-LLM on a single RTX A6000 — Docker GPU passthrough, KV cache, prefill/decode, Dynamo, and parallelism strategies.May 2026