Company · Research
What we arebuilding toward.
Our research is the part of the work that does not fit in one engagement: the questions every customer raises and no single project can answer. We work on them between projects and put the results back into the next one.
A model that answers questions is useful. A system that finishes the work, reads the record, writes back to it and can be checked afterwards, is what a company actually needs. Our research is the distance between the two.
Why we do it
Research that ships.
We are not a laboratory. Every direction on this page came out of an engagement where the available tools were not good enough, and every result goes back into one. The test of a piece of research here is not a paper; it is whether the next customer's workflow runs on its own.
That keeps the work honest in a way that matters for AI. The systems we build have to be right on a customer's real records, not on a benchmark, and they have to stay right as those records change. Most of what we work on is the gap between a demonstration and a system that runs unattended.
Where the work is general enough to be useful outside our own engagements, we write it up and publish it here, with the method attached so it can be checked.
Directions
Six things we keep coming back to.
Each is a question we meet on most engagements. Each maps to a service where the current answer is already in use.
- 01
Agents that finish work
Long tasks across several systems, with the permissions of one person, and a record of every step that can be reversed.
- 02
Models that learn a company
Training and tuning on a company's own records so the model knows its products, its customers and its vocabulary.
- 03
Evaluation that holds up
Golden sets that predict production behaviour, and re-scoring that finds drift before a customer does.
- 04
Reading documents whole
Invoices, contracts, forms and scans read into structured records, with every field traceable to where it came from.
- 05
Voice that does the job
Speech in and out that books, confirms and transcribes in the languages a company's customers actually use.
- 06
Operations that run unattended
Whole operations — an office, a floor, a warehouse — run by systems that watch themselves and call a person only for the exceptions.
How we publish
What a result has to carry.
- 01
The method, not just the number
Anything we publish comes with how it was measured, on what kind of data, and what it was compared against, so the result can be reproduced or argued with.
- 02
No customer data
Research is done on our own records, on public data, or on data a customer has agreed in writing to contribute. Nothing about a customer is published without that agreement.
- 03
Failures included
A direction that did not work is written up the same way as one that did. The next engagement needs to know both.
- 04
Open where it can be
Where a result is general and does not depend on a customer's systems, we release the code and the evaluation set with it.
Questions
Asked about the research.
Where are the papers?
There is nothing listed yet. This page describes the directions we work on and how we publish; results will appear here as they are written up, with the method attached.
Do you train models on customer data for research?
No. Customer data stays in the customer's environment and is used only for that customer's work. Research uses our own records, public data, or data a customer has explicitly agreed in writing to contribute.
Can we collaborate on a research question?
If it is a question we meet on engagements and you have the data and the patience to answer it properly, yes. Write to us through the contact page and say which direction it belongs to.
Is research a separate service?
No. It is the part of the work that does not fit inside one engagement. If you need something built, start with the services; the research shows up in how they are built.
Start
Tell us what your teamstill does by hand.
We map the workflow with you and say plainly what can run on its own, what should not, and what it takes.