Cutting the AI bill for engineering teams
A service for teams overspending on AI. I look at where the money goes, move work to cheaper models and cache the answers that repeat.
Token prices fell sharply in 2026, and yet enterprise AI bills climbed anyway: usage grew faster than the unit cost dropped. I packaged that pain into a FinOps service for engineering teams.
The problem
Teams that live on coding assistants and production LLM inference watch the bill grow without knowing what drives it. It isn't a price-per-token problem; it's a usage-architecture problem.
The approach
The core is a waste-driver taxonomy — what the audit hunts:
- Model over-selection → a model-tiering policy + routing (small model where it's enough).
- No prompt caching → restructure for cache hits (stable prefix).
- Runaway agentic loops → max-turn caps and cheap sub-models for sub-agents.
- No visibility → per-dev/team quotas, alerts, and a dashboard.
- Cloud-only for everything → hybrid routing to local models for eligible tasks.
On top of that I built the offering (audit → reduction sprint → fractional FinOps) and a landing page.
The result
My own bill, running a hybrid local↔cloud stack on a consumer GPU, is the case study: every percentage I report is measured, never invented.