Work out whether you need one
Often you do not. We start by testing whether retrieval, better prompting, or a smaller hosted model already clears your bar — and we tell you when the honest answer is that fine-tuning would be a waste of your budget.
In-house LLM development
Your data, your weights,
your infrastructure.
Sometimes the constraint is regulatory, sometimes it is cost at volume, and sometimes it is simply that a general model has never seen your domain. We build and serve models that live inside your perimeter.
The work
What we actually do.
Build the dataset
The model is the easy part. We assemble, clean, and label the training data, and we build the held-out set that tells you whether any of it worked.
Fine-tune and distil
Adapt an open-weight model to your domain, then distil it down until it fits the latency and cost envelope you actually have to hit in production.
Serve it properly
Quantisation, batching, GPU sizing, autoscaling, and a rollback path. A model that only runs on the researcher's laptop is not a deployment.
What you get
- feasibility assessment
- training + eval datasets
- fine-tuned / distilled model
- inference service
- benchmarks vs baseline
What we work in
- Llama
- Mistral
- Qwen
- vLLM
- PyTorch
- LoRA / QLoRA
Also from us
The rest of what we do.
The demo worked. Then it met real inputs.
Agentic workflow development
Multi-step agents that do real work, with the guardrails and evals that let you put them in front of customers.
Read more →Someone on your team does this every morning, and it takes an hour.Agentic automation
Agents pointed at the repetitive internal work your team does by hand and resents.
Read more →You need three hundred calibrated questions, and they have to be original.Technical content services
Assessment and learning content for platforms that put problems in front of engineers.
Read more →Start with one workflow.
The fastest way to find out whether this fits is to pick something narrow and build it. Tell us what you have in mind.