Custom LLM development
Custom LLM training and deployment.
We take proven language models and make them yours: fine-tuned on your own data, deployed on your cloud or on your premises, connected to your documents, and tested hard before anyone in the business relies on them. We adapt existing models to your work. We do not claim to build foundation models from scratch, and we do not need to.
Services
Six pieces of work, one accountable team.
Most engagements need two or three of these. Some need all six. Each is scoped and priced on its own, so you pay for the work the problem actually requires.
Fine-tuning on your data
We adapt an existing model to your terminology, formats and judgement calls using your own examples: past drafts, resolved tickets, approved reports. The result behaves like someone who has worked in your business, not a generalist guessing.
Private deployment
Open-weight models deployed on your cloud account or on-premise hardware, so prompts, documents and outputs stay inside infrastructure you control. Engineered for your data, owned by you, running in your stack.
Retrieval over company documents
Retrieval-augmented systems that answer from your policies, contracts, manuals and records, and show the passage each answer came from. When the documents do not contain the answer, the system says so.
Evaluation and red-teaming
Before go-live, the system is tested against your real cases and against deliberate attempts to break it: invented figures, false citations, instructions hidden in uploaded documents, requests it should refuse. You see the results before you decide.
Model routing and cost control
Not every request needs the largest model. We route each task to the model that does it well at the lowest sensible cost, with failover between providers, and report what each application spends.
Ongoing monitoring
Models drift as your documents, products and customers change. We watch quality and cost after launch, re-run the evaluation when something changes, and retrain when the evidence says it is needed.
Engagement
How a custom model gets built.
01 · Scoping
A call to define the task, the users and what good looks like, then a short written proposal. Fixed-fee where the scope is clear, milestone-based where it is not.
02 · Data readiness
We review the data you have, what it covers, what it is missing and what cannot be used. If the data will not support the model you want, you hear that before the build, not after.
03 · Build
Fine-tuning, retrieval, or both, built against the agreed task. Findings shared as they surface, not saved for a final deck.
04 · Evaluation
Tested on your real cases and red-teamed against failure. You see where it is strong, where it is weak, and what it will refuse to do.
05 · Deployment
Deployed on your cloud, on-premise or through MNB Intelligence, with access keys, logging and human approval on anything irreversible.
06 · Handover
Documentation, evaluation sets and runbooks your internal team can maintain, with ongoing monitoring if you want us to keep watching.
Who it is for
Built for organisations whose knowledge is the advantage.
Enterprises with sensitive data
Listed companies, regulated businesses and groups whose documents cannot be pasted into a public chatbot, and who need the model inside their own walls.
Growing businesses with a document-heavy core
Firms where the work is drafting, checking, answering and reconciling, and where a model that knows the house style saves real hours every week.
Teams already spending on AI
Organisations with several AI tools and no single view of what they cost or how they behave, who want one governed route to the models instead.
What you own at the end
- The fine-tuned model, deployed where you chose, under the licence of the model it was built on
- The evaluation set it was tested against
- Documentation and runbooks your team can maintain
- A readable log of what it did and what it cost
What we will not do
- Claim we trained a foundation model from scratch
- Put a model live before it has passed an agreed evaluation
- Put an agent in front of money, tax invoices or customer commitments without a human approval step
- Quote an accuracy number we have not measured on your data
Questions
Before you call.
Do you build foundation models from scratch?
No. We fine-tune and adapt existing models, open-weight and commercial, on your data and for your task. Building a foundation model from scratch is rarely what a business problem needs, and we do not claim to do it.
What is the difference between fine-tuning and retrieval?
Fine-tuning changes how a model behaves: its terminology, formats and judgement. Retrieval changes what it knows at the moment of answering, by reading from your documents. Most business systems need retrieval; some also need fine-tuning. Scoping decides which.
Can the model run on our own infrastructure?
Yes. Open-weight models can be deployed on your cloud account or on-premise hardware, so prompts, documents and outputs stay inside infrastructure you control. Where a commercial model is the better fit, it can be routed through MNB Intelligence or a private deployment of it.
Which models do you use?
Whichever does the task well at a sensible cost. We are not tied to one provider, and model routing lets different tasks use different models behind one API.
How do you know the model is ready?
It is evaluated against your real cases and red-teamed against failure before go-live: invented figures, false citations, instructions hidden in uploaded documents, and requests it should refuse. You see the results and decide.
How is a custom LLM engagement priced?
With a scoping call and a short written proposal. Fixed-fee where the scope is clear, milestone-based where it is not. There is no retainer requirement.
Scope a custom model.
A 30-minute call with the founder, then a written proposal that says what the model will do, what data it needs and how it will be tested.