Fine-tune models on your data, with the evidence to prove it helped.
When prompting and retrieval are not enough, we build the dataset, run the training and measure the result against the model you started with. You get a model, a model card and the numbers.
Outcomes we deliver.
Dataset design and curation
We define what the model needs to learn, source and clean the examples, and document provenance so you know exactly what went in.
Annotation and human review
Guidelines, labelling workflows and quality checks, with your domain experts reviewing the examples that matter most.
Supervised and preference tuning
Supervised fine-tuning for task behaviour, preference tuning for tone and safety, on open-weight or provider-hosted models.
Domain adaptation
Teaching a model your vocabulary, formats and edge cases: clinical shorthand, product codes, regulatory language.
Deployment-ready artifacts
Quantised or hosted checkpoints, serving configuration and the inference code to run them in your environment.
Before-and-after evaluation
Every training run is measured against the baseline on a held-out set you have seen, so the improvement is real and repeatable.
Deliverables you own.
- Data card describing sources, licensing and processing
- Annotation guidelines and quality metrics
- Training recipe and reproducible pipeline
- Model card with evaluation results and known limitations
- Serving setup and hand-over session
Four steps, measured at each one.
Baseline
We measure the current model (or prompt) on a held-out set so there is a number to beat.
Data
Design, collect, clean and label the training set with your experts in the loop.
Train
Run and compare training configurations in your environment, tracking every experiment.
Evaluate and ship
Report the gains and the regressions, then package the model for serving.
What we will and won’t do.
What we work with.
Chosen per engagement, never by default. No vendor logos because we are not paid by any of them.
Who this is for.
- Teams whose prompts have hit a ceiling on accuracy or consistency
- Companies that need a model running inside their own environment
- Products where latency and cost per request make smaller tuned models attractive
Common questions about ai model training.
When does fine-tuning beat prompting or retrieval?
When the task needs consistent format or behaviour across thousands of calls, when the domain language is far from general text, or when you need a smaller model for cost, latency or residency reasons. If prompting and retrieval would do the job, we say so in the discovery sprint.
How much data do we need?
Less than most people expect for behaviour and format, often hundreds to low thousands of high-quality examples. Teaching new knowledge needs more, and is often better solved with retrieval. The baseline step tells us how far the current model is from the target, which drives the data plan.
Who owns the model?
You do, subject to the licence of the base model. We hand over weights, pipelines and documentation. We do not keep copies of your data or your model after the engagement ends.
Where does training run?
In your cloud account or on your hardware by default. If you have no GPU capacity we can provision rented capacity under your account so the data never leaves your control.
Do you also train people?
Yes, as part of hand-over. We run working sessions with your engineers so they can rerun the pipeline, extend the dataset and evaluate the next version without us.
Tell us what you're trying to ship.
Thirty minutes with the founder. We'll tell you honestly whether we're the right team, and what it would cost.