AI systems / inference / research
Inference engines, model adaptation and agent infrastructure. I take ideas through implementation, evaluation and production, then publish what the measurements actually show.
Software engineer at AWS ElastiCache
Maintainer of Valkey GLIDE
01 / Selected work
Four projects that show how I work: close to the model, close to the hardware and honest about the baseline.
Inference engine / Rust + CUDA
An OpenAI-compatible engine for Blackwell GPUs, built around the constraints that matter in serving: kernels, quantization, speculative decoding and cache behavior. Text, image and video models run through one serving path.
Model adaptation / GRPO + LoRA
A small zero-initialized gate inside Qwen3.5's Gated DeltaNet layers changes what length-penalized reinforcement learning teaches the model. The result is shorter reasoning with higher measured accuracy, not a prompt trick.
Serving systems / learned prediction
A 421M return-time predictor decides where an idle session's KV cache should wait: GPU, CPU, disk or recomputation. The placement rule uses costs the engine measures on itself, with no fitted constants.
Agent infrastructure / open source
Twenty-five public repositories for making coding agents useful beyond a demo: deterministic CLI selection, local model routing, configuration validation, desktop control and isolated workspaces across more than 50 agent platforms.
02 / Open source
I work on infrastructure other engineers depend on, from official clients to the review systems around them.
All projects03 / Research
Every study states the protocol, the comparison and the caveat. A result is only useful when you can see what would overturn it.
Research index ↗A zero-training recipe reaches 1.8× to 2.7× end-to-end speedups. Pruned, low-rank and distilled alternatives lose 10 to 19 acceptance points under distribution shift.
Read paper and failed arms ↗Remove half the routed experts, heal the surviving function and quantize. The public record includes the community mask that collapsed at the screening gate.
Read method and artifacts ↗05 / Contact
I work on inference, model adaptation, evaluation and the infrastructure around AI systems.