01
An evaluation harness
A versioned test set from your real cases, scored on every change. You can read it, extend it and run it without us.
- Service A
- N° 0042 / Applied AI_
- 6 to 14 weeks
- From 38k EUR
Applied AI that can show its working.
Retrieval, agents and decision support that earn their place in a workflow, shipped with a test set and a cost ceiling.
01 / What you get
Four things you keep.
02
The feature, behind a flag
Retrieval, a drafting assistant or an agent with a short list of tools. Reversible, rate-limited and logged.
03
Guardrails and a cost ceiling
Input and output checks, a refusal path that sounds human, and a per-user budget that fails closed.
04
A watch dashboard
Quality, latency and spend on one screen, with drift alerts routed to the rota that owns them.
02 / The loop
Measured at every turn.
01 / Frame
Write the failure down.
We turn the brief into a test set of fifty to two hundred real cases, with the answer a careful human would give.
02 / Measure
Score every change.
Each prompt, model or retrieval tweak is scored against the set. No change merges on a vibe.
03 / Ship
Release behind a flag.
Ten percent of traffic, then half, with a kill switch and a spend ceiling that fails closed.
04 / Watch
Catch the drift early.
Live traffic is sampled back into the set. When quality slips, the rota hears about it before your customers do.
03 / Engagements
Pick a shape.
2 weeks
Sprint
A prototype on your data and a go or no-go memo.
- From
- 38k EUR
- Team
- 3 people
8 to 14 weeks
Build
A production feature with evals, guardrails and a rota.
- From
- 96k EUR
- Team
- 4 to 5 people
Quarterly
Embed
Two of ours inside your team, shipping alongside yours.
- From
- 52k EUR / quarter
- Team
- 2 people
Prices are invented placeholders.
04 / Questions
Asked often.
Which models do you use?
Whichever scores best on your test set at a cost you accept. We route through a vendor-neutral gateway, so changing provider is a config change and your prompts and evals stay yours.
Can it run on our own infrastructure?
Yes. We have shipped on a private cloud and fully on-premise. It costs more in setup and we will tell you by how much in the discovery week.
Who owns the evaluation set and the prompts?
You do, from the first commit. They live in your repository and they work without us.
What happens when the model gets worse?
The watch dashboard samples live traffic back into the test set. A drop past the agreed threshold pages the rota, and the flag lets you fall back to the last good configuration.
Do you sign off on regulated work?
We build to your compliance team's checklist and keep an audit trail. We do not certify anything on your behalf.