Original writing from gogoai.dev on SRE, AIOps, Platform, DevOps, and Cloud Ops.
Why teams running real enterprise AI workloads are quietly migrating from vector databases to knowledge graphs — and what it means for cost, latency, and model size.
Frontier API access is a structurally expensive, strategically fragile bet at scale. The winning architecture minimizes tokens consumed per unit of business value.
Weights, KV cache, and engine overhead all share VRAM. A practical guide to sizing a GPU for multi-user LLM serving — with the formula and worked examples.
Every LLM request splits into prefill (compute-bound, sets TTFT) and decode (memory-bound, sets TPOT). Understanding both is how you control latency and cost.
RAG, tool calling, and agents are the three building blocks. Default to the simplest architecture that works — a 5-step agent costs 5x and fails 5x more.
A single-page reference for serving LLMs in production: vocabulary, sizing formulas, GPU and model tables, essential vLLM flags, and cost levers ranked by impact.
24 real production LLM mistakes — from deploying Ollama for multi-user to unbounded conversation history — each with the fix. Read it before you ship.