Technical articles on engineering, research, and leadership.
A systematic, non-destructive workflow for diagnosing why a Kubernetes service is misbehaving. Covers pod-level diagnostics, networking investigation, node pressure analysis, centralized log correlation, and structured root cause synthesis.
How llama.cpp's prompt caching mechanism works, why the default 8 GiB allocation caused KV cache eviction errors on my server, and what the parameter actually controls versus what most guides get wrong.
An empirical evaluation of local LLM inference on Apple M5 Max with 128 GB unified memory. This study benchmarks eight models across prompt processing and token generation workloads, demonstrating that Mixture-of-Experts architectures achieve 3 to 6x higher throughput than dense models on memory-bandwidth-bound hardware, and presents a production deployment architecture for agentic coding workflows.
A comprehensive reference guide to large language model terminology, covering architecture, quantization, fine-tuning, inference, and the naming conventions you need to navigate the local LLM ecosystem.