AI agents and LLM integration · Build
AI agents that prepare each case for a person to decide
Where a wrong action costs money or trust, the model prepares each case and a person decides. At a German financial institution, operators confirmed every case.
EngagementHourly · remote
Hours9–18 CET · 4 h overlap
StackClaude Code · TensorFlow · Node.js
Proof2 case studies
The service in 4 parts
01 · Who it’s for
A team that wants a model or an agent inside a real workflow, where a wrong action costs money or trust and someone has to answer for it.
02 · What we’ve built
- For a German financial institution’s remote identity checks, the model made a confident recommendation in about 98% of cases, and an operator confirmed every case. The 98% measures only how often the model was sure of itself.Case study 10 →
- A language model in daily engineering work: Claude Code writes integrations, components and tests for a US marketplace, and nothing it writes is accepted until the team’s automated checks pass.Case study 03 →
- An afternoon of research in which an AI was confidently wrong six times, and every error leaned toward a tidier answer.Blog post →
The model in the identity work was a computer-vision model, not a language model. The arrangement around it applies to a language model as well: the model suggests, an operator confirms and records a reason, and the decision can later be traced to that operator.
03 · How we work
01 · CASE FILEEach decision is saved with its reason, the person who made it and the history of the case, so it still holds up when someone reviews it years later.
02 · PERMISSIONS PER ACTIONPermissions are granted action by action. In the identity system, an operator who lacked the permission for an action handed the case to a colleague who had it. An agent gets its permissions the same way.
04 · Stack
Claude CodeTensorFlowNode.jsAWSper-action permissionsdecision records
Proof
Where we’ve done this before
Case study 10 · Identity98%For a German financial institution’s remote identity checks, the model made a confident recommendation in about 98% of cases, and an operator confirmed every case. The 98% measures only how often the model was sure of itself.Read the case study →Case study 03 · Platforms~50%A language model in daily engineering work: Claude Code writes integrations, components and tests for a US marketplace, and nothing it writes is accepted until the team’s automated checks pass.Read the case study →WritingThe biggest fool: six confident AI errors in one afternoonAn afternoon of research in which an AI was confidently wrong six times, and every error leaned toward a tidier answer.Read the post →
FAQ
Questions about this service
For drafts and suggestions, yes. Anything that moves money, changes a record or reaches a customer waits for a person to approve it, because in regulated work someone has to answer for the decision.
Confidence and correctness are separate measurements. On the identity project we tracked how often the model was confident and never established how often it was right. Next time we measure both.
Confident answers that rest on assumptions nobody stated. We ask for the assumptions before the answer, and then ask what would have to be true for the result to be wrong.
Next step
Describe the problem and we’ll come back with three options.
The first call takes 30 minutes, and you’re under no obligation.