Generative AI that cites its sources
Document intelligence, company search and copilots, tested against your own examples before they reach your team.
- Answers traceable to source
- Tested on your own data
- Model choice stays open
What we build with generative AI
Each one ends in a record or a person, never in a chat window alone.
Document intelligence
Invoices, contracts, forms and reports become structured data, with low-confidence fields sent to a person.
Company knowledge search
RAG across manuals, policies and records, respecting who is allowed to see what.
Team copilots
Assistants inside the tools your sales, finance and operations teams already use.
Customer support assistants
Answers from your documentation, and a clean handover when it runs out.
Drafting and summarising
Quotes, replies, reports and meeting notes written from your data, for a person to send.
Evaluation and rescue
Already have a GenAI pilot that disappoints? We measure it and find out why.
A good answer shows its source
A sample of how it should behave, including when it should refuse.
What is the warranty on the X200 pump?
24 months from delivery, excluding seals.
warranty-policy.pdf, section 3.2
What will the X200 cost next year?
I can’t find a future price in the documents you can access.
Handed to the sales team
Five tests before launch
A demo proves it can work. These prove it works on your data.
| Test | The question it answers |
|---|---|
| Accuracy | Is it right on examples your team has checked? |
| Citations | Does each source actually say what the answer claims? |
| Permissions | Can one user retrieve another department’s documents? |
| Prompt injection | Can text inside a document change what the model does? |
| Cost per task | What does one finished task cost at expected volume? |
Questions we get asked first
Which model do you use?
We choose per task by testing candidates on your own examples. That can mean a hosted model such as Claude, GPT or Gemini, or an open-weight model where cost or data rules demand it. The system is built so that the model can be swapped later.
Will our data train a public model?
That depends on the provider and the contract, so we check the terms for your chosen service before any data moves. Where the rules are strict, we can design for private deployment or an open-weight model on infrastructure you control.
What is RAG?
Retrieval-augmented generation finds relevant passages from your own documents and gives them to the model along with the question. Done properly, it lets answers cite their sources and refuse when the documents don’t contain an answer.
Can generative AI connect to our existing systems?
Yes. It can read from your ERP, CRM, document stores and databases through their APIs or the Model Context Protocol, and write drafts back for review.
How do you stop hallucinations?
You can’t remove them entirely. You reduce them with retrieval, citation checks and evaluation sets, and contain them by keeping a person in the loop where a wrong answer costs money.
What do people keep retyping? Let’s start there.
Name the documents, where they arrive and what happens to them next.
Start the conversation