02 · LLM applications
Assistants that know your business.
We build LLM applications — internal assistants, drafting copilots and natural-language interfaces — that answer from your content, follow your rules, and behave the same way on the thousandth request as on the first.
What we build
Four shapes of LLM application.
Internal assistants
Answer staff questions from policies, procedures and systems — with citations and links to act.
Drafting copilots
First drafts of reports, letters, responses and summaries in your format and tone, ready for a person to finalise.
Natural-language data
Ask questions of databases and dashboards in plain English, with generated queries checked before they run.
Customer-facing chat
Help on your website or app that stays on topic, hands over to people cleanly and never invents policy.
Engineering
The difference between a demo and a dependable tool.
Anyone can wire a chat box to a model. The engineering is in making it accurate, safe, affordable and maintainable.
- Grounding and citations — answers come from retrieved sources, and show them.
- Structured outputs — responses validated against schemas so downstream systems can trust them.
- Prompt management — versioned prompts and system instructions, reviewed like code.
- Evaluation suites — agreed test questions scored automatically on every change.
- Guardrails — topic limits, input and output filtering, and safe refusal behaviour.
- Cost and latency controls — caching, model routing and token budgets, monitored per feature.
FAQ
LLM application questions.
Can the assistant connect to our single sign-on?
Yes. We typically integrate with your identity provider so the assistant knows who is asking and only uses content that person is allowed to see.
What does it cost to run?
It depends on volume, model choice and context size. We estimate running costs during scoping and build in caching, routing and monitoring to keep them predictable.
How do you measure whether it’s good enough?
We build a test set of real questions with your subject-matter experts, agree a pass mark for accuracy and tone, and score every release against it.
Thinking about an internal assistant?
Tell us who would use it and what it should know. We’ll outline how we’d ground it and measure it.