Blog
All the articles I've written.
-
Inference Optimization Mini-Benchmark: FP16 vs Q8_0 vs Q4_K_M
Hands-on benchmark of quantization levels on a 1.5B model: decode throughput, TPOT, and memory footprint compared across FP16, Q8_0, and Q4_K_M.
-
Serving the Agentic Inference Era — Optimization Strategy (Slides)
Slide deck on optimizing LLM inference for the agentic era: quantization, KV cache, batching, and bandwidth-bound decoding.
-
I Ran vLLM on a Mac Mini With No GPU — Here's Everything I Learned About Inference
A complete, beginner-friendly guide to vLLM: what inference actually is, how to build vLLM from source on an Apple Silicon Mac with no GPU, every command explained, the three errors I hit and fixed, the flags that matter, and real throughput numbers from my living room.
-
OpenCode Doesn't Have an Auto Mode. So I Built One with vLLM Semantic Router
Cursor picks the right model for you automatically. OpenCode doesn't - yet. Here's a complete, production-shaped guide to adding intelligent auto model selection to OpenCode (or any OpenAI-compatible agent) using vLLM Semantic Router and AgentGateway.