See the spend before finance does. Then cut it.
A large share of most AI bills is waste: the biggest model for every call, prompts that repeat the same context, no caching and no budget per feature. We measure first, then cut, and only ship a change when the evals prove quality did not move.
Cut AI costs
Model routing, caching, prompt compression and token budgets, with observability built in.
What we do
Cost audit
Where every token goes: per feature, per customer, per model. Two weeks of traces tell the truth better than the invoice.
Model routing
Small models for easy tasks, large models for hard ones. The eval suite decides which is which, not a guess.
Caching and compression
Prompt caching, deduplication, shorter contexts and batch processing for anything that does not need an answer now.
Budgets and alerts
Token budgets per feature or customer, cost attribution in Langfuse and an alert before the number becomes a surprise.
How it works
- 1
Measure
We instrument your calls with traces and let two weeks of real traffic build the cost map.
- 2
Build the safety net
An eval suite that scores the outputs you care about, so every optimization has a before and an after.
- 3
Cut, one change at a time
Routing, caching, compression, budgets. Each change ships only if the evals pass.
- 4
Keep it that way
Dashboards, budgets and a monthly review so the bill does not creep back.
What you get
- Cost map by feature, customer and model
- Routing layer and cache in your codebase
- Eval suite with before and after results
- Cost dashboard, budgets and alerts
Good fit if
- A monthly LLM bill that is already noticeable or grows with users
- Traffic in production, so there is something to measure
- A team willing to let evals decide
Not a fit if
- A prototype without traffic yet (build it right from the start instead)
- A single call per day where the bill is not the problem
Stack we use
Questions
Will quality drop?
Only changes that pass the eval suite ship, and you see the before and after for every one of them.
Which providers do you work with?
Any. We are vendor-neutral and often mix providers, with open models through Ollama where it makes sense.
Do we need Langfuse?
Any tracing works. We default to Langfuse because it is open source and you can host it yourself.