Now

I help companies choose, adapt and deploy open models on their own hardware or cloud account. We compare options on the customer's workload and budget, then agree the deliverables, handover and support before the work starts.

• Model and hardware assessmentKnow what to run before buying hardware: candidate comparison, workload baseline, cost recommendation.
• Deployment and optimizationFit the model and serving engine to hardware within the budget: installed configuration, measurements, operating instructions.
• Fine-tuning and evaluationMeasure the quality gap, then train where the evidence supports it: evaluation set, candidate comparison, deployment artifacts.

The lab's models are on Hugging Face: NVFP4 builds of GLM, DeepSeek, MiMo, Qwen and DictaLM, MTP draft heads in GGUF, and Phonon-2 in ONNX.

Discuss a project

Negative-eigenvalue gates for hybrid models. The retrofit works at 4B and 9B: +48 and +43 points on state tracking at 4x the training length, over LoRA alone. I'm scaling it to Qwen3.8-27B, where the current run trains with thinking on and measures accuracy and reasoning tokens per task.

Layby, released. KV cache placement for idle LLM sessions by when they come back. On a slow disk with Qwen3-8B in vLLM, p95 time to first token at 0.59x of an LRU CPU tier with the median unchanged, and 0.49x of the better of two LMCache settings. Code, model and benchmark are open; the paper goes to arXiv next.

Finished changes waiting on maintainers.

• sglang SM120 sparse MLA: zero-initialize the 64-token page-split scratch #42440
• sglang SM120 DeepSeek V4 decode: read FlashInfer's dispatch envelope #42445
• sglang Tune long FP8 asymmetric prefill on RTX PRO 6000 #42451
• sglang HiCache: size the hybrid Mamba host tier, with per-pool metrics #42443
• sglang HiCache: keep the device-only SWA match gate #42446
• sglang Keep a non-linear NextN layer out of the Mamba layer list #42441
• sglang Opt-in strict DSML parsing for DeepSeek V3.2 and V4 tool calls #42442
• sglang Qwen3Coder: recover object parameters with an unescaped quote #42448
• sglang Fix MiMo packed-module mappings for ModelOpt policies #42450
• sglang Defer serialized CUTLASS NVFP4 scale layouts until loading #42452
• sglang Redact credentials in server_args logs, /server_info and launch_command #42444
• sglang Router: Valkey backend for sgl-kv-indexer #42447
• sglang Router: Valkey event log, worker replay and indexer failover #42449
• valkey Add sync-from-replica with a delayed primary switch #3724
• llama.cpp Imatrix-aware NVFP4 quantization with scale search, and the NVFP4 file type #25153
• llm-compressor Fix activation caching for dataclasses that validate their fields #3144
• LMCache Support binary buffer objects in the local disk backend #3285
• LMCache Handle local disk read failures safely #3292

All my open pull requests to other projects, live on GitHub.

Valkey GLIDE and valkey-skills: review, releases and support. The agent-sh tools, mainly agnix and parlar. My day job is in-memory data systems at AWS ElastiCache.