Service / AI
Fine-tuned LLMs that know your house style.
Fine-tuning and distilling open language models so they follow your format, tone and policies, run faster and cost less per request than a general model.
[ 01 ]The short answer
What is LLM Fine-Tuning and Distillation?
LLM fine-tuning adapts an open language model, such as Llama, Qwen or Mistral, to your task by training it on examples of the output you want. Distillation goes further: a large teacher model labels your data and a small student learns to match it. The result follows your formats and rules more reliably, answers faster and costs less per request. Fine-tuning teaches behaviour, not facts, so we pair it with retrieval when answers must be current.
What you get
- Fine-tuned weights or LoRA adapters
- Curated training set and data card
- Evaluation report against base and frontier models
- Serving configuration
- Retraining recipe
[ 02 ]What we build
The pieces, built to last.
- 01
Supervised fine-tunes
LoRA, QLoRA or full fine-tunes on curated examples, so the model writes in your structure every time.
- 02
Distilled students
A frontier teacher labels your traffic and a small model learns to match it, for most of the quality at a fraction of the cost.
- 03
Preference tuning
DPO and related methods that teach the model which of two answers your experts prefer.
- 04
Dataset curation
De-duplication, PII scrubbing, balancing and hard-example mining, because the dataset decides the result.
- 05
Side-by-side evaluation
Blind comparisons against the base model and a frontier model on your test set, per segment.
- 06
Efficient serving
Quantised weights, multi-LoRA serving and batching, so several tuned adapters share one GPU.
[ 03 ]Where it pays off
Where it earns its keep.
- 01
Structured extraction at scale
Turn millions of documents into JSON that validates against your schema, at a cost that works in volume.
- 02
Brand and policy compliance
Drafts that follow tone, disclaimers and regulatory wording without a long prompt on every call.
- 03
Domain language
Clinical, legal or financial phrasing that general models get almost right, which is not right enough.
- 04
On-device and edge
Small models that run on a laptop, phone or private server where data cannot leave.
[ 04 ]How we deliver
Four steps, one team, start to finish.
The same people scope it, build it and stay after launch. Timings are typical for a first release; we agree yours after the deep dive.
- 01Week 1
Find the decision
We pick the decision or task worth automating, measure how it is done today and agree what better means in numbers.
- 02Week 1 to 2
Build the eval first
A labelled test set from your own data comes before any prompt or training run, so every change is scored, not argued.
- 03Weeks 2 to 10
Build with guardrails
Models suggest, code decides. Permissions, injection checks, human override and audit logs are part of the first version.
- 04Ongoing
Ship and improve
We launch to real users, watch cost, latency and quality per request, and feed failures back into the eval set.
[ 05 ]Stack
Tools we reach for.
Chosen per project, and we work in yours. Nothing here is a lock-in: you own the code and the accounts.
- Llama
- Qwen
- Mistral
- Gemma
- Hugging Face TRL
- Axolotl
- Unsloth
- PEFT
- vLLM
- SGLang
- OpenAI and Anthropic APIs as teachers
[ 06 ]Related work
Built by the same team.
[ 07 ]Further reading
Problems we found, and how we would solve them.
- Models3 min read
Fine-tune, retrieve or prompt? A decision guide
ProblemTeams pick fine-tuning, RAG or custom training by reputation rather than by what the problem needs, and spend months on the wrong one.ProposalDecide by what must change: prompts for instructions, retrieval for knowledge that changes or must be cited, fine-tuning for behaviour, format and cost, and custom models only where the task or data is truly unusual; always prove it on an evaluation set.Read the proposal - Cost3 min read
Your AI bill is a routing problem
ProblemAI applications send every request to a frontier model with a long prompt, so cost and latency grow with usage while most requests never needed that capability.ProposalRoute each request to the cheapest model that passes your evaluation for that kind of request, cache shared prompt prefixes, batch what can wait, and distil a small model for the highest-volume task.Read the proposal
[ 08 ]FAQ
Questions buyers ask.
Use retrieval for knowledge that changes or must be cited, and fine-tuning for behaviour: format, tone, policy and speed. Many production systems use both. Our guide to the choice is linked below.
For format and style, a few hundred high-quality examples can be enough. For distillation we usually label tens of thousands of real inputs with the teacher model.
Check the provider's terms first: some restrict training competing models on their outputs. We review the terms with you and use open teacher models where needed.
[ 09 ]More AI services
- Custom AI Model DevelopmentA model trained on your data, owned by you.
- AI Agent DevelopmentAgents that finish the job, inside the rules.
- AI Workflow AutomationHours of back-office work, done in minutes.
- RAG DevelopmentAnswers from your documents, with citations.
- AI Memory and Knowledge GraphsAI that remembers what your organisation learned.
- MCP Server DevelopmentYour product, usable by every AI assistant.
- Voice AI AgentsEvery call answered, qualified and logged.
- Document AI and Intelligent Document ProcessingPaperwork turned into clean, checked data.
- Private LLM DeploymentAI on your infrastructure, data never leaves.
- LLMOps, Evaluation and GuardrailsKnow your AI works before customers find out.
- Predictive ML and ForecastingDecisions made on what will happen next.
One company at a time