Example · Cost
Swap a large model for a small one
Fine-tune a small open model on one high-volume task so it runs faster and cheaper, with an eval that shows whether quality holds.
Fine-tuning & custom models
Models trained on your data, for your exact task.
When a general-purpose model is too slow, too expensive or not accurate enough, we fine-tune or train one that is. We measure it against what you use today, so the numbers back the decision to switch.
What we build
Smaller, sharper models that you own.
We start from your task and your data, not from a favourite model. If fine-tuning isn't the right tool, we'll tell you in the audit.
- Fine-tuning open-weight models such as Llama, Mistral, Qwen and Gemma
- Fine-tuning through provider APIs, when that fits better
- Custom classifiers, extractors and ranking models
- Building the dataset: cleaning, labelling and synthetic data
- Evaluation suites that compare against your current model
- Deployment to your cloud, your own servers or the edge
Example projects
Where it pays off.
These are examples of first projects, not client case studies. We prove it on your own data first.
Example · Accuracy
Domain-tuned extractor
A model trained on your own documents to pull out the fields that generic models keep getting wrong.
Example · Brand
On-brand writer
Tuned on your approved content so drafts sound like you and follow your rules.
How it runs
Measured before you switch.
-
01Day 0 · 30 min
Audit
A free call about the problem, your data and your constraints. You leave with a ranked plan.
-
022 weeks
Prototype
A first fine-tuned model, scored next to your current one on your own test set.
-
03Before go-live
UAT & sign-off
Your team tests it on real cases and we fix what they find. Nothing goes live until you sign off.
-
04To ~day 90
Ship
Hosting, monitoring, drift alerts and a retraining pipeline your team can run.
-
05Ongoing
Scale
We extend it, tune it and train your team, then hand it over when you're ready.
Built to be trusted
Your model, your data, your call.
You own the weights
The model, the data pipeline and the evals all live in your accounts.
Your data trains only your model
It is never used for a shared model or for another client.
Numbers before switching
We recommend a switch only when your own evals show the new model is better.
FAQ
Straight answers.
Anything else? Ask us on the audit call, or just email.
How much data do we need?
It depends on the task. Some narrow tasks improve with a few hundred good examples, while others need thousands. We check what you have during the audit and tell you honestly whether fine-tuning is the right tool.
Fine-tuning or RAG?
RAG is better when the model has to look up facts that change. Fine-tuning is better when it has to learn a format, a style or a narrow skill. Many systems use both, and we'll tell you which fits.
Where does the model run?
Wherever you choose: your cloud account, your own servers, a managed inference provider, or on the device itself for small models.
Who owns the result?
You do: the weights, the training code, the datasets and the evals.
Bring us the task your current model gets wrong.
A free 30-minute call. You leave with a ranked plan, whether or not you work with us.