Model Routing Doctrine — Cut Your LLM Bill 30–60%
The opinionated routing doctrine I use to run a 24/7 multi-agent fleet without bleeding money. If your team is using one frontier model for everything, this single document typically cuts LLM spend 30–60% on real workloads.Written for engineers and operators who are past the demo and now have to make the bill stop growing.WHAT'S INSIDEThe three-tier framework: mechanical, generalist, high-judgment — with concrete examples for each.The routing assignments: which model tier handles which task, and why. Local models for Tier 1, mid-tier APIs for Tier 2, frontier models for Tier 3.The one-page decision tree. Tape it to your monitor.The cost-per-outcome formula — the only metric that actually matters — with the cost-modeling spreadsheet schema you can build in Google Sheets in 20 minutes.The measurement template: exactly what to log on every LLM call so you can see savings on real traffic.Four worked examples from production: content generation (77% savings), code review automation (80%), document data extraction (97%), customer support (77%).Seven anti-patterns that quietly inflate your bill, including the most expensive sentence in AI engineering.A day-by-day plan for applying the doctrine in week one.WHO IT'S FOREngineering leads watching their LLM line item climb past comfortable. Founders who built on the easiest API and now need to scale economics. Anyone who has ever typed "Claude Opus" into a config file because they didn't have time to think about it.WHO WROTE ITAdrian G. — AI Agent Infrastructure Engineer based in Basel, Switzerland. Operator of a 24/7 multi-agent production fleet running across Linux, macOS, and Windows. The doctrine in this PDF is the same one keeping that fleet under budget.FORMATPDF, ~25 pages. Drop-in for any team running LLMs in production.
Get it → adriangercak.gumroad.com