AI build and integration for companies
We build the AI toolyour business actually needs.
A business need, a tool designed and shipped into production. Eight published builds, from a radiology practice to a food processing plant, each with its duration and its measured results. When new software is not the answer, we plug AI into the systems already in place.
Six written questions first, then real slots on the same page. Free, 30 minutes.
What the eight published builds add up to
Four numbers, and where to recount them.
These numbers are the sum or the average of the eight builds published on the Proof page. Each one can be recounted there, record by record, from the durations and volumes on display.
- 8
- published builds, across 8 sectorsFrom three radiology sites to two food processing lines.
- 446
- weeks of building in totalThat is 56 weeks on average per build, from scoping to production.
- 5,905
- documents, files and requests handled each working daySum of the five builds that carry a counted document flow, with weekly, monthly and quarterly rates brought back to a working day.
- 2023
- the year of the first AI tool we put into productionA document chain, with mandatory human review before signature. The pattern has not changed.
No organisation named, no logo: our agreements forbid it. Each build is designated by its sector and its size, which lets you judge how comparable it is to your own situation.
The industry changes, the method does not.
We build custom AI tools. How they get built does not depend on the sector; what depends on it is the work they support.
Start from the daily gesture
We look at what someone actually does every day, not at a list of features.
Write the scope before building
What gets built, what does not, and the criterion that says it is finished. Written before the first line of code, and binding afterwards.
Build for production
The tool ships into the real system, with its access, its data and its failures, not into a demo.
Measure, or abstain in writing
When the data is not enough to conclude, the abstention is written down and its reason is named. A number that cannot be defended does not get published.
Integrate rather than add one more piece of software
A new tool is not always the right answer. When the systems already in place hold up, we plug AI into them instead of replacing them. Same team, same method, different scope.
One build, in three lines
A radiology practice wanted to produce its reports faster. We built a tool where the radiologist dictates the diagnosis to an agent, the agent writes the report in the practice's own template, and the radiologist reviews and signs in the same screen. 14 weeks of building. Median delay between dictation and signed report: from 19 h 40 to 4 h 15.
This build is described by its sector and its size, without naming the organisation. The other seven are on the Proof page, in the same shape.
Three ways in, a written scope every time.
You can stop after scoping. No work starts without a written scope, named deliverables and a completion criterion.
Scoping
The way in. We qualify the need, then we write the scope, the deliverables and the completion criterion.
Custom build
We design and deliver the tool that was asked for, all the way into production.
Integration into what exists
We plug AI into the systems and the processes already in place.
No public pricing. We sell a scope, named deliverables, a duration and a completion criterion, not a listed package.
How an estimate is put togetherEight builds, and two products open to inspection
What we have built.
The eight delivered builds are published by sector and by size, each with its duration and two measured results. They cannot show their code: the agreements forbid it. The two products below can, because our teams publish them, and they carry the full measurement with its initial failure.
Minealyze
The technical problem
Detect which paying user is churning on high-volume behavioral telemetry, explain why in natural language, act, then measure the effect against a control group or refuse to conclude.
QuestFable
The technical problem
Turn a natural language sentence into a valid structured artifact, deployed into a third-party system with an undocumented proprietary format, under a constrained grammar.
- The eight builds, with their durations and their figures, are on the Proof page
- The screens of both products sit behind authentication
- The measurements, their dates and their limits are on the Proof page
The measurements, their threshold, and their instrument.
The rest of the site, and what each page holds.
About
The company since 2023, its actual registration, its five build roles, and how a project is run.
What we do not do
The requests we turn down, each with its written reason, and pointers to who does it better.
Security
What is in place, what does not exist, and the headers actually returned with the date they were captured.
Website terms of use
The website terms of use, in their dated version.
A qualification call, then a named service.
There is no “let's discuss your needs”. There are six written questions, then real slots on the same page.
Six written questions first, then real slots on the same page. Free, 30 minutes.