AI that works in your real business
An AI demo can be built in a weekend. A system that handles your permission rules, your 200,000 documents, your compliance requirements and your most demanding users needs proper architecture.
That is the work we do.
Where to go next
Five pages, one for each kind of AI work we do.
AI Assistants (RAG)
An assistant that knows your company
AI Agents & Automation
Software that takes over repetitive work
AI for Your Existing Product
Build the AI feature your customers keep asking about
How We Build with AI
Where AI does the work, and where humans do
Free AI Opportunity Audit
Where AI will pay off, and where it won't
Start here
The free AI Opportunity Audit
Many companies come to us with a solution already in mind, aimed at the wrong problem. So we start here, with a free 60-minute session.
You walk us through how your team works day to day. We look for places where someone reads information in one place and types it into another, or answers the same question for the tenth time that week.
What you get, in writing, within three days
- The three processes where AI would pay for itself fastest
- A rough hours-saved estimate for each
- What each would cost to build, and how long it would take
- The things you should not use AI for, and what to do instead
- Which one to do first, and why
The audit is free and there is no obligation. If you take the report to another developer, that is fine. Either way you will have a much better brief than you started with.
What we build
AI assistants that know your company
Built on your policies, contracts, manuals, product data and past tickets. Staff or customers ask in plain language and get an answer with the exact source. If there is no source, the assistant says it doesn't know and offers to pass the question to a person. Retrieval is permission-aware, so users only get answers from documents they were already allowed to open.
AI agents that finish a task
Software that reads emails, pulls out the data, updates your systems, chases approvals and escalates anything that is stuck. When it isn't confident enough, it asks a person.
Document reading and extraction
Invoices, purchase orders, contracts, forms, medical records, CVs and shipping documents are read, extracted, checked and filed. Anything below the confidence threshold goes to a review queue for a person to check.
AI inside your existing product
You already have software, and your customers are asking about AI. We add the feature to your product, in your stack and under your brand, without rewriting what already works.
Search that understands meaning
Someone searches "refund after 30 days" and finds the policy that says "returns outside the standard window."
Private and on-premise models
When your data cannot leave your building, we run open-weight models on your own infrastructure. This costs more and performs slightly worse, but for some organisations it is the right trade-off.
How we decide whether AI is the right answer
We include this because honest advice matters more to us than selling AI.
AI is the right tool when
The input is messy and unstructured, the rules are fuzzy and full of exceptions, a person currently reads something and makes a judgement, or the volume is too high for people to keep up with.
AI is the wrong tool when
The rules are clear and fixed, in which case a rules engine is cheaper, faster and always right. The data is already structured, so a database query will do. Being wrong even 2% of the time is unacceptable and nobody is checking. Or the real problem is messy data, which needs fixing first or you simply automate the mess.
We regularly turn down AI projects for these reasons
It may cost us work, but it saves you a six-month lesson.
Five safeguards built into every AI system
01
A source on every answer
If the system cannot cite a source, it doesn't answer.
02
Permission-aware retrieval
The system can only find what that specific user is allowed to see.
03
A measured accuracy number
We build a test set of real questions with known correct answers and report the score before launch, including the areas where it is weak.
04
A human review queue
Anything below the confidence threshold waits for a person. Anything involving money, health, employment or legal risk always does.
05
Cost controls
Hard spending caps, per-user limits and alerts, so a runaway loop never shows up on next month's bill.
How an AI project runs
- 1
Free 60-minute audit
We find the opportunities and rank them.
- 2
Two-week proof of concept
A working prototype on your real documents and questions, with an honest accuracy report. If it isn't good enough to be useful, we tell you, and you have spent two weeks finding out instead of six months.
- 3
Production build, 4 to 10 weeks
Security, permissions, monitoring, cost caps, human review queues and integration with the tools your team already uses, such as Teams, Slack, your portal or your CRM.
- 4
Tuning in the first month live
Real users always ask things nobody predicted, so we tune retrieval and prompts against the questions people actually ask.
Priced on scope
You receive a written proposal covering the build and the expected monthly model usage, with hard spending caps in place.
Our AI stack
- AI
- Azure AI FoundrySemantic KernelOpenAI APILangChainPostgreSQL with pgvectorAzure AI SearchOpen-weight models on your own hardware.NET and Python
Questions people ask
Will it make things up?
Any language model can. We reduce the risk in three ways: it can only answer from your documents, it must show its source, and it declines when it has nothing reliable to point to. We then measure it against a fixed set of questions and show you the score before launch. Be wary of any vendor who says their system never makes things up.
Does our data get used to train someone's model?
No. The enterprise API tiers we use do not train on your data. If you need more assurance, we can run models inside your own cloud or on your own servers.
How much data do we need?
Less than you might think. A few hundred documents is enough to prove whether the idea works.
How long before we see something real?
You get a working proof of concept on your own data within two weeks.
What does it cost to run every month?
It depends on usage volume. We estimate the monthly cost before you build and set hard caps so the bill never surprises you.
We tried an AI project last year and it failed. Why would this be different?
Tell us what happened on the call. Most failures come down to one of four causes: a chatbot with no specific job, no sources so nobody trusted it, no accuracy testing before launch, or messy underlying data. All four can be avoided before you spend serious money.
Not sure yet what you need?
Book a 30-minute call with a senior architect and tell us about the problem. We will explain what it would take to build, roughly what it would cost and whether custom software is the right answer at all.