Build
Custom-trained AI models
Most people who ask for a fine-tuned model need good retrieval instead. We will tell you which you are.
What is in scope
- Assessment of whether fine-tuning is actually the right answer
- Training data preparation, which is the bulk of the effort and the bulk of the cost
- Fine-tuning and evaluation against a held-out set
- Deployment, usually self-hosted
- Retraining as the data shifts
How the engagement runs
- Baseline first. We measure a retrieval-based approach on your task. Frequently it is sufficient and the engagement stops here, which is the cheapest possible outcome for you.
- Data assessment. Fine-tuning needs consistent, labelled, reasonably plentiful examples. Most businesses find they do not have them yet.
- Train and evaluate. Against held-out real cases, with the baseline as the comparison.
- Deploy and monitor for drift.
What determines the price
Dominated by data preparation, not compute. If your examples are already clean and labelled this is affordable; if they have to be assembled, that is the project.
When this is the wrong service
Stated plainly, because an engagement that should not have happened costs us more than it costs you.
- You want the model to know your documents. That is retrieval, and it is cheaper, faster and easier to update.
- You have fewer than a few hundred consistent examples.
- The task changes frequently. A fine-tune is a snapshot.
Bring us the process that keeps breaking
A short call is usually enough to tell you whether this is the right service.
