What problem does it solve? Local LLM runtimes like Ollama, llama.cpp, LM Studio, and vLLM are frequently misconfigured — models overflow VRAM into slow CPU offload, quantization levels degrade output quality, context lengths exhaust memory, and routing logic leaks sensitive data to cloud endpoints. This Skill produces a concrete configuration review with calculated VRAM requirements, throughput baselines, and privacy boundary verification instead of generic optimization advice. ## Core Features & Use Cases - Hardware and Model Fit Analysis: Calculates VRAM requirements using weights plus KV cache formulas, identifies CPU offload penalties, and validates quantization levels against task types. - Privacy and Routing Audit: Verifies that local-only data never reaches cloud endpoints including fallback conditions, checks telemetry settings, API binding, and outbound network traffic during inference. - Throughput Benchmarking: Measures time-to-first-token, tokens per second, and peak VRAM against reference benchmarks to detect configuration problems. - Use Case: Before deploying Ollama to process medical records on a machine with a 24 GB GPU, run this review to confirm the quantized model fits in VRAM, the router enforces privacy classification before capability routing, and no telemetry sends prompt content externally. ## Quick Start Review my local Ollama setup running Llama 3 8B on a 16 GB VRAM GPU and check the configuration, privacy boundaries, and throughput before I use it with sensitive data.
