How we work
One method, whether we build it or you do.
Four steps, the same for an agent we build and for a feature your team built. Each step ends with something you can read, and the price is fixed before the second one starts.
01Design
What happens
A kickoff call. We map the job, the users, the risks and what a good result looks like. For an agent, we write down what it may do, what needs a person, and what it must never do. For evaluation work, we write the risk map and the rubric.
What you get
A design sheet or a risk map and rubric, and a fixed-price proposal with the scope written down.
02Build
What happens
We build in your stack and your repository, with the test set written alongside the code. Weekly check-ins, and working software early rather than a reveal at the end. For evaluation work, this is where the test set and the gates get built.
What you get
Code and tests in your repository, with documentation your engineers can follow.
03Prove
What happens
Every change is scored against the test set: quality, groundedness, safety, tone and the cases that matter to your business. Safety checks run in an agreed scope. We walk you through the results in a readout.
What you get
An evaluation report with scores, failures, evidence and fixes, and a go or no-go recommendation.
04Run
What happens
After launch, every prompt, model, retrieval or tool change gets an evaluation run within two business days. A monthly safety sweep, a monthly report, and a quarterly review of what to do next.
What you get
A release gate in your pipeline, a monthly report your leadership can read in five minutes, and a team that knows your product.
The rules we work by
- Fixed scope and a fixed price, agreed before we start. Scope changes are written down, never absorbed quietly.
- Your stack, your repository, your tools or open source. You own the deliverables.
- Evidence over opinion: every finding in a report links to a test case you can rerun.
- The release decision is yours. We give you a recommendation and the evidence behind it.
- Client accounts only, with multi-factor authentication and the least access that does the job. Access is removed within 24 hours when someone leaves.
- No production personal data on our devices, ever.
- An NDA and a data-processing agreement before the work starts.
- Honest about limits: no one can promise bug-free software, and we do not.
What we need from you
One kickoff call
About 45 minutes, to agree the job, the risks and what good looks like.
A test environment
Accounts you provide, with test data you approve.
One contact
Someone who can answer a question within a day.
Access to the repository
For builds and release gates: your repository and your CI.
What you get, on paper
The three documents that change hands most often. Each one is written to be read in minutes.
- Example
Evaluation report
A score for every release, with the failures, the evidence and the fixes.
- Example
Agent design sheet
What the agent may do, where a person approves, and what gets logged.
- Example
Release check
A gate in your pipeline that stops a release when a check fails.
Tell us what you are building.
One email is enough. You will hear back from the founder, not a sales team.
Founding clients get a lower price on their first project.