AI tools for your team's work.
We help you work out which jobs are worth giving to language models. To do that we study your processes and build a pilot. Once it has proved itself in real work, we train your team.
Cursor, Claude, GPT and Gemini are tools we use every day, so we speak from experience. In a consultation we go through your processes and tell you what AI still cannot do. Then we pick the tasks where its value can be put to the test.
Start with a defined implementation scope
How we can help
AI strategy
We study your processes and pick out the tasks that are worth automating, with an estimate of the return on each. The result is a plan for the next 6–12 months.
RAG systems
We build search across your company's documents. The language model answers from the material it retrieves and says where the information came from.
AI agents
Agents take over some of the work in support, sales and internal operations. We connect the tools they need, set up what they should remember and decide in advance what an agent is allowed to do.
Team training
Hands-on sessions with Claude, GPT and Cursor. During them we write prompts and templates with your staff for the tasks they do every day.
Fine-tuning
Sometimes an off-the-shelf model is not up to the task. We then check whether fine-tuning would help, using SFT, DPO and distillation and evaluating the quality. The updated model is rolled out in stages.
Evaluation & safety
We test answers against a set of queries and log errors and made-up claims. Separately, we test the defences against stray instructions that can reach the model through external data.
Toolkit
We choose models and tools on four counts: answer quality, cost, speed and where the data has to be hosted. We work with open models and with vendors' hosted services alike.
How we run a project
Initial assessment
Takes 1–2 weeks. We study your processes, shortlist tasks for AI and estimate the return. Whether to go further after the assessment is your decision.
Pilot project
We take one real task, in customer support for example, and test a solution on it over 1–2 months. We launch it and measure the result.
Ongoing support
From three months. We maintain and develop your AI systems, train staff and help with new tasks as they come up.
Practical guides and deep dives
Before we start
Where should LLM adoption start in a company?
Start with a short audit. It takes 1–2 weeks: we study your processes, find the tasks that suit AI and estimate the likely return. Whether to carry on after that is up to you.
Which models do you use?
Claude Sonnet 5.5 and Opus 5.5, GPT-6 Astra, Gemini 3.1 Pro and 3.8 Flash, and the open models Llama 4, Qwen3.8, DeepSeek V4, Kimi K2.6 and MiniMax M3. Which one we pick depends on the project: answer quality, cost and latency all count, and so does whether the model has to run on your own servers.
How is RAG different from fine-tuning?
RAG leaves the model itself unchanged: the system retrieves relevant documents and passes them to the model along with the question. That suits information which changes often. Fine-tuning alters the model's parameters and can help it keep to a particular style or response format. The two can be combined.
How do you measure AI quality?
We measure quality on a set of real user queries. Before each release we run the set again and compare the answers with the previous version, noting errors and made-up claims. The results show whether the new version is fit to release.
What would you like AI to help with?
The consultation takes an hour and costs nothing. We will talk through your processes, decide where to begin and show you how to weigh the costs against the benefits.