AI, built like software.
The AI part of your product, and the automation of work inside your business. Measured against a written set of questions before it ships, and measured again every time it changes.
The part of a product that a demonstration never shows: what happens on the thousandth request.
AI inside your product
Getting something impressive working once is an afternoon. Getting it to hold up when real people use it, on real data, at a cost you can predict, is the actual job. That second part is what we do.
- Answers from your own data
- Your documents, records and history made searchable by meaning rather than by keyword, so the system answers from what your business actually holds instead of from what a model remembers.
- Doing things, not just saying things
- Giving the model real actions it can take in your system, with limits on what it is allowed to do and a record of every action it took.
- Choosing what runs where
- Different work needs different models, and the cheapest one that passes is the right one. We measure that rather than guess, and we keep the choice swappable.
- The part that survives production
- Cost per request, how long an answer takes, what happens when the model is wrong, slow or down, and logging detailed enough that a bad answer can be reproduced and fixed.
No product involved. Work your staff currently do by hand, done by a machine, with a person on the exceptions.
Automation inside your business
Paperwork that reads itself
Invoices, forms and documents turned into data your systems can use, with the ones it is unsure about sent to a person instead of guessed.
The questions answered twice a day
The things customers and staff ask constantly, answered from your own material, with anything unusual handed to a person.
Assistants over your own files
Staff asking plain questions of company material they would otherwise have to go looking through, with access limited to what each person is allowed to see.
The difference between a feature and a demonstration is that somebody measured one of them.
How we know it works
Step 1
Before anything is built, we write down the questions it has to get right, together with you. That set is the target.
Step 2
We run it against that set, so there is a number to compare against rather than an impression of how it feels today.
Step 3
We run it again on every change, because the usual way these systems fail is quietly, months later, when something upstream moved.
Step 4
If it cannot be measured, it does not ship. A demonstration that works once is not a feature.
We would rather lose the work than build something expensive that a report would have done.
When we say it is the wrong tool
A great deal of what is being sold as AI right now is something cheaper wearing a costume. If that is what your problem turns out to be, we will tell you, and quote you for the cheaper thing.
When it is a database query
A great deal of what gets sold as AI is a question with one correct answer sitting in a table. That wants a report, not a model, and it costs a fraction as much to run.
When it is a form or a rule
If the decision is the same every time given the same inputs, write the rule. It will be faster, cheaper, and you will be able to explain it to a regulator.
When being wrong is not survivable
Some decisions cannot carry a system that is usually right. Where that is the case we will say so, and where it still makes sense we put a person on the end of it.
The same terms as everything else we build. Nothing here runs only on our hardware.
What you own
- Your data stays yours and is not used to train anything for anyone else.
- The prompts, the evaluation set and the logs are yours, in your repositories.
- It runs on your accounts, and nothing in it depends on hosting that only we can operate.
Bring the problem rather than the technology. We will tell you if it needs any of this.
Tell us what you want it to do
Describe the job in plain words: what comes in, what should come out, and how you would know it got it right. We will tell you whether this is work for a model or for something simpler, and what it would cost.