How we work

Check first, then build.

AI projects rarely fail because of the model. They fail because of data, interfaces, or a task nobody made measurable. That is why every project starts with a check on real data and ends with a system measured against criteria agreed beforehand.

Three steps

From use case to running system

  • 01

    Check

    In the Readiness Check we examine the use case on your real data: state of the data, interfaces, deployment path, measurability. It ends with a written verdict — build or don’t. A reasoned “don’t build this” is a legitimate result.

  • 02

    Build

    We build in weekly steps, each with working software. What “done” means is fixed beforehand: thresholds on a frozen test dataset. The evaluation runs before acceptance and as a regression test after every change.

  • 03

    Hand over

    We put the system into operation on your infrastructure and hand over source code, documentation and reproducible deployment into your repositories. You run it yourself afterwards, or we keep looking after it.

The deliverable

A report you can base a budget decision on

A document with six chapters that management can read and IT can recalculate.

The check as a decision gate 01 KICKOFF 02 ANALYSIS 03 REPORT 04 DECISION Kickoff workshop2 hours, remoteor on site Analysisof your real data,systems and processes Written report 01Summary & verdict 02Use-case analysis 03Data & systems check 04Architecture & operating costs 05Risks & GDPR 06Roadmap & offer plus a 90-minute results session Gobuild in weekly steps No-goreasoned stop beforefurther effort is spent Duration: 1–2 weeks
Both outcomes are a result: a go starts the build, a no-go stops the project before further effort is spent.
Acceptance

“Done” is a measurement, not a feeling

An AI system does not return the same answer every time. So before the build we agree not on an output but on a threshold over a frozen test dataset. A criteria sheet can look like this:

acceptance-criteria.yamlexample
test_dataset:
  size: 200 reference questions from the department
  state: frozen before project start
criteria:
  source_correct: >= 90 %          # answer names the right source
  answer_without_evidence: 0      # rather “not found” than guessed
  access_rights_violated: 0
  response_time_p95: <= 4 s
measurement: before acceptance and after every update

The values are an example. Criteria and thresholds are set together for each project.

Questions about a project?

Lukas Friedrich answers questions about how we work personally.