Software · Data and AI
Generative AI
Applications that draft, summarise and answer questions using language models grounded in your own content.
Language models are good at first drafts and summaries, and poor at knowing when they are wrong. We build generative AI tools around that: answers grounded in your documents with citations, outputs checked against your rules, and a person reviewing anything sent to a customer or regulator. Quality is measured on your own examples before rollout.
What we build
- Knowledge assistants that answer staff questions from policies, manuals and past tickets
- Drafting tools for proposals, product descriptions and replies in your house style
- Meeting, call and case summaries written straight into your CRM or case system
- Retrieval pipelines that index documents with permissions, so people see only what they may
- Evaluation sets and dashboards that track answer quality, cost and latency per release
- Private deployments of open-weight models where data must stay on your infrastructure
DemoInteractive · sample data
Try it live.
A working sample with made-up data. It runs in your browser; nothing is sent anywhere.
Atlas Ask me about company policies. I answer from the HR and finance handbooks and cite the section.
How many casual leaves do I get?
Can I claim a client dinner?
What it does well
Grounded with citations
Answers link to the paragraphs they came from, so readers can check them in seconds.
Permission-aware retrieval
Retrieval respects access rights, so an answer never draws on files the user cannot open.
Style and rules
Tone, banned claims and required disclaimers are set in prompts and checked on output.
Measured quality
Each release is scored on a fixed set of your real questions before it reaches users.
Cost controls
Caching, smaller models for simple tasks and usage limits keep model spend predictable.
Human sign-off
Drafts meant for customers, contracts or regulators are reviewed and edited before sending.
Questions
Is our data used to train the model?
We use business API terms under which providers state they do not train on your data, and we confirm the current terms for your case in writing. Where rules require it, we run open-weight models inside your own cloud account so data never leaves it.
Which model should we use?
There is no single best model. We test a few on your actual tasks and compare quality, speed, cost and where data is processed. Often the answer is a mix: a stronger model for hard drafting and a smaller, cheaper one for classification and short answers.
How do you stop it from making things up?
We cannot remove the risk entirely, so we reduce and contain it. Answers must come from retrieved sources, missing sources lead to a refusal, outputs are checked against rules, and anything with legal, financial or medical weight is reviewed by a person.