Free PDF · ~12 pages
Cost-Aware LLM Deployment
Three patterns that keep the monthly bill under $200/mo while still using frontier models where they earn it. Cost-guard fallback, per-agent budgets, and the Grafana panels that surface breaches before the invoice does.
What’s inside
- The cost-guard pattern, in copy-paste form. A 20-line wrapper around the model router that swaps every paid-API call to a local fallback when an agent crosses 80% of its monthly budget.
- Per-agent budgets that scale with task value. A worked `models.yaml` showing budget allocation across Researcher, Hero Writer, Bulk Writer, and QC. Why "local-only" is a real config choice, not a fallback.
- Local fallback that actually works. Ollama vs. vLLM trade-offs, when to fall back per-call vs. per-agent, and the operational gotchas that bite a fresh laptop.
- Three Grafana panels that earn their pixel. Spend per agent vs. budget; token rate per agent; cost-per-campaign distribution. The exact dashboard JSON ships with the PDF.
Who it’s for
- Founders shipping LLM features and watching the bill grow faster than usage
- Engineering leads picking a model-routing strategy before the first prod deploy
- SREs / platform owners adding LLM observability to an existing Grafana stack