Process

How an engagement works

We run every project in clear stages with written agreements, defined milestones, and acceptance criteria you approve. Here is exactly what to expect.

Predictable by design

Custom AI work has a reputation for open-ended timelines and unclear results. We counter that with structure. Each stage has a defined output, a review point, and a decision: continue, adjust, or stop. You never pay for months of work without seeing evidence of progress.

Four connected stations along a path representing project stages

1. First call and NDA

We start with a free 30 to 45 minute scoping call. If you would like to share confidential details, we sign a mutual non-disclosure agreement before the call or right after it. We are happy to use your NDA template or ours.

2. Discovery and statement of work

For most business projects, we recommend a short paid Discovery phase. We interview stakeholders, inspect a sample of your data, review the license terms of candidate base models, and produce a written plan. The plan becomes a statement of work listing deliverables, milestones, acceptance criteria, timeline, and price.

3. Data handling

Data is transferred through an encrypted channel that you approve, or we work entirely inside your cloud account if you prefer. Access is limited to the engineers assigned to your project. We use your data only for your project, never for any other client, and delete it at the end of the engagement or on request, with written confirmation. Details are on our Responsible AI page.

4. Milestones

A typical training or distillation project has three to five milestones, for example:

  1. Data prepared and evaluation set agreed.
  2. Baseline results for the starting model.
  3. First trained candidate and evaluation report.
  4. Refined model meeting acceptance criteria.
  5. Handover and deployment.

After each milestone, we hold a review meeting and share a written summary. You can redirect priorities based on what the results show.

5. Acceptance criteria

Before training begins, we agree in writing what "done" means. Criteria are measurable wherever possible, such as a target score on the evaluation set, a maximum response time on specified hardware, or a pass rate in expert review. If a target turns out to be unreachable with the available data, we tell you as early as possible and discuss options, including stopping the project.

6. Delivery

Final deliverables typically include model weights, an evaluation report, training scripts, and a deployment guide. We walk your team through everything in a handover session.

7. Support

Every project includes 30 days of follow-up questions by email. After that, optional monthly support plans cover monitoring, help with issues, and scheduled retraining. See deployment and support.

Not sure where to start?

Book a free scoping call. We will look at your use case, your data, and your constraints, and tell you honestly whether a custom model makes sense.

Book a free scoping call