A model trained for your task
When tuning is justified, we train an open-weight model on your approved proprietary data. You receive the model weights or adapters agreed in the project scope.
AI models trained on your data.Deployed Privately.
We find where your model falls short, shape your data, and measure whether tuning helps. You approve the budget and see the results before deploying. If training isn’t the answer, we’ll tell you.
Industry perspective · Model ownership
“Companies need to turn their workflows, domain knowledge, and accumulated judgment into AI systems that improve with each use.”
Satya Nadella
Microsoft
In his article A frontier without an ecosystem is not stable, Nadella argues that companies should retain control of their knowledge and use their own evaluations to measure whether AI improves business results.
Read the article on XRevolut Cut AI Coding Costs by 90% Using a Self-Hosted Open-Source Model!
David Abitbol
Revolut
Speaking at AI Engineer NYC, Abitbol said: Open source is now a priority at Revolut, and our economics are favorable.
The Register reports Revolut cut its AI coding inference costs by roughly 90 percent by self-hosting an open model.
[Deliverables]
What you receive.
Every engagement starts with a task and data review, a baseline evaluation, and a scoped recommendation. If we proceed with tuning, these are the deliverables we agree with you.
Measured results against your baseline
Your evaluation report compares quality, cost, and speed on the same held-out test set. It shows improvements, regressions, tradeoffs, and whether deployment is justified. Lower inference cost is a measured goal; improvements are not guaranteed.
A private API on dedicated GPUs
If results justify deployment and you approve it, we deliver an authenticated inference API on dedicated GPU capacity, with integration documentation. We agree capacity, access controls, data handling, and operating responsibilities before deployment.
You own your data and tuned model, subject to the base model’s license. We agree who operates and maintains the endpoints.
[Models delivered]
Three models. Three business tasks.
Educational content
Irish education
Extract book metadata from educational content into structured fields.
MeasureField-level accuracy and valid JSON on held-out documents.
MethodTask-specific eval harness before any tuning recommendation.
Study-level extraction
Cancer research
Extract study-level information from cancer research documents for researchers to review.
MeasureStudy-level F1 and reviewer-aligned gold set on excluded test studies.
MethodGroup-safe splits so near-duplicate studies never cross train and test.
Resume authenticity
Hiring
Flag resume claims for authenticity review. People investigate the flags and make hiring decisions.
MeasurePrecision on flagged claims vs human review; cost per reviewed resume.
MethodBaseline on generic models, then SFT only when flags improve without excess noise.
[Models & methods]
The right model starts with the task.
We select an open-weight base and tuning method around your data, success criteria, and budget. Model eligibility and licensing are confirmed before training.
[What we measure]
Does the tuned model improve your task?
We compare the tuned model with your current model on the same test examples. You see what improved and what got worse before deciding to deploy.
- Accuracy
- Compare correct answers, missing fields, and errors against your current model.
- Cost
- Measure inference costs, retries, and human review costs per correctly completed task.
- Speed
- Measure response time and the time people spend checking and correcting outputs.
Before training, we agree which metric must improve and which limits must hold. A cheaper model is not useful if it makes too many errors.
[How it works]
Start with the problem. Measure before you train.
You bring domain expertise. We handle the ML engineering.
Your team defines success and reviews results. We prepare the data, tune the model, and measure what changed.
[FAQ]
Questions before you start.
How much data do we need?
It depends on the task and the quality of your examples. We check whether the examples cover your task, have consistent answers, and leave enough separate examples for testing.
What will it cost, and how long will it take?
We propose a budget and timeline after reviewing your task and data. Training starts only after you approve the plan.
How do you decide whether the tuned model is better?
We agree success criteria before training, then compare your current model and the tuned model on the same held-out test set, excluded from training and model selection. We measure quality, cost, and speed and show the tradeoffs. You review the report and approve any production deployment.
What if tuning does not help?
You receive the results and our recommendation: improve the data, try another approach, or keep your current model. We do not guarantee an improvement.
Where will our data and model run?
The deployment offer uses dedicated GPU capacity with an authenticated inference API. Dedicated GPUs do not by themselves place the model inside your own VPC or private network. We agree hosting location, access controls, and data handling before you share a dataset.
Who owns the data and model?
You own your data and your tuned model. The base model retains its original license, which governs permitted use and redistribution. Ownership and handoff terms are recorded in the project agreement.
Do we need an in-house ML team?
No. We handle data preparation, tuning, and evaluation. Your team provides a task owner who can judge examples, define success, and review results, plus an engineering contact to connect the API and coordinate deployment.
[Start a project]
Tell us the task.
Describe what the model should do, what it gets wrong today, and what data you have. An engineer will reply within one business day.
We start by reviewing your task and available data, then scope a baseline evaluation and recommend whether tuning is worth pursuing.
Describe your data here; do not include sensitive records. We agree scope and budget before training starts.