Engineers inside the product team
A product team wants AI inside the product and does not yet have the practice around it. Our engineers join their repository and their planning, ship the first feature with them and leave everything in their codebase.
The problem
The company wants the product itself to help the user: draft, classify or look something up inside the workflow, in their own interface. A separate chatbot on the side would miss the point.
The team can ship product code. What is missing is the AI engineering practice around it: evaluations checked in with the code, traces, cost limits and a clear rule for when the model may call a tool or change data. Hiring a full-time specialist would take longer than the feature can wait.
What we build
Our engineers join the team. They work in the client’s repository, attend the client’s planning and open pull requests like everyone else.
The first goal is one feature in production, with the harness around it:
- The model call lives in their codebase, behind the same review as the rest of the product.
- An evaluation set of real product cases runs in continuous integration.
- Traces show what the feature did, what it cost and where it failed.
- A person, or a hard rule in code, confirms actions that write to customer data.
After that, the client’s engineers take more of the changes. We review, extend the evaluation set and add the next feature only when the first one is stable.
How it runs
- We are embedded in their team, in the hours our working days overlap.
- Their stack, their accounts, their data. We work where the code already is.
- A weekly demo of what shipped, and a short note of what the evaluation set showed.
- Length and team size are agreed on the first call.
What the client gets
- Senior engineers from the first week.
- A feature in production, with tests and an evaluation set.
- A team that has learned the practice by building it, and can extend it after we leave.
- Everything we wrote, in their repository, under their ownership.
The result
One AI feature is live inside the product. The next feature starts from the harness, the traces and the evaluation set already in the repository.