AI that survives contact with real data.
A chatbot that answers from your own material. An assistant inside your app that does the thing rather than pointing at where to find it. Search that understands a question. And agents that take a repeatable process from one end to the other, so nobody has to sit in the middle of it.
What we build.
AI-powered products at the front, automation at the back. Assistants and chatbots people talk to, and agents that pick up the repetitive work nobody wants to keep doing by hand.
- 01A chatbot on your site or inside your product, answering from your own documents and handing the conversation to a person when it should.
- 02An assistant inside the app: it drafts the message, finds the record and fills the form.
- 03AI in a mobile app: read a document with the camera, dictate a note, ask a question and get an answer from your own data.
- 04Search and extraction: statements, invoices, contracts and forms turned into fields your systems can act on.
- 05Agents that run a repeatable process end to end, calling the same systems a person would have opened.
The model is the easy part. What takes the time is the data around it, who is allowed to see what, and what happens when the answer is wrong, because sometimes it will be. That is where the work goes: the retries, the logging, the way to replay a bad run, and the point where a person steps in.
What you get.
- 01An audit trail for every answerEvery prompt, the documents the model was allowed to touch, and the tools it called, written down and kept. When someone has to explain why the system said that, it is on the record rather than reconstructed afterwards.
- 02Permissions that holdThe model sees what the signed-in person is allowed to see and nothing else, enforced in your system rather than asked for in a prompt.
- 03A set of real examplesYour own cases with the right answers attached, so a change to the prompt or the model is checked against them rather than judged by feel.
- 04A person on the decisions that countAnything that moves money, writes to a customer record or sends a message waits for a human yes, unless you decide otherwise in writing.
- 05Running cost you can seeUsage reported per feature, so what the AI costs to run each month is a number you agreed rather than a surprise on an invoice.
- 06The code and the prompts, in your nameThe system, the prompts and the examples are yours, in your repository, and portable to another model or another supplier.
The demo is not the hard part.
Anyone can show a model answering well on a question it was shown in advance. What decides whether the thing is worth having is the ten thousandth run: the malformed PDF, the customer who phrases it differently, the tool that times out halfway. We have spent years building software where a failed request costs someone real money, and we build these the same way, then stay answerable for them after launch.
Questions, answered.
Almost never. Training from scratch costs far more than the problem is usually worth. We use a hosted model, put your own data around it so the answers come from your material, and only look at fine-tuning if a measurable gap is still there afterwards.
Wherever you decide, and it is written down before anything is built. The major hosted models can be used without your inputs being kept or trained on. Where that is not enough, the work runs against a model in your own cloud account instead. If a document must never leave your network, we design around that rather than asking you to accept the risk.
It will, so the design starts there. Anything that moves money or reaches a customer waits for approval. Everything else is logged with the inputs it saw, so a run can be replayed and the cause found rather than argued about.
That is most of this work. Adding to a system that is already live and cannot be stopped is what we do on every other service too. An AI feature reads from what exists, writes back through the interfaces that exist, and goes to one team before it goes to everybody.
Sometimes it isn’t. If the process changes shape every time it runs, or it happens twenty times a week, a short script or a better form will beat an agent and cost less to keep. We will tell you that before you have committed a budget rather than after.
Send us the questions people actually ask.
The ones your support inbox gets every week, or the job you would hand to a new starter and the documents they would need to do it. You’ll get a straight read on whether AI helps here, what it would take, and where it would go wrong.