Line B · AI consulting
Scoped AI work, with a date on it.
For software firms getting AI features into production. We scope in one or two sessions, deliver named artifacts against that scope, and hand the work to your team in a state they can build on.
Deliverables
You’ll be running it without us.
The measure of a good engagement is what your team can do after it. So the scope names the artifacts and the date, and the work is done when those land, with your engineers able to run, change and extend it from day one.
In practice that means code they can read, integrations wired into your stack rather than sitting beside it, and a walkthrough that hands over the reasoning and not only the result.
And if the honest answer is that you don’t need this yet, you’ll get that too. We’d rather be the people you call next year.
Where we’re most useful
- ✓There’s a real workflow to improve, with someone who owns it
- ✓Your engineers will take the result and keep going
- ✓You want it shipped, not surveyed
Good problems to bring
- A process buried in email, spreadsheets and copy-paste
- A first AI feature your team hasn’t had time to start
- A prototype that works in a demo and nowhere else
- A domain that’s too specific for a generic tool
Who does the AI work
The people who know where these systems break.
Not a vendor who read the documentation last month. Someone who works on the failure modes for a living, and writes them down.
Our partner studio
Iliano is also the AI engineer at Faraday AI Studio in Rome, an OpenAI Select Partner that puts AI into production for companies across Italy and Europe, and whose founder is one of n8n's ambassadors. That is the automation stack Runset builds booking flows on, so the overlap is practical rather than decorative.
Runset delivers its own work. What the studio gives it is a bench to test a design against before it reaches you, and a standard for what counts as finished.
Iliano Fasolino runs the AI line at Runset and publishes research in association with the University of Milan, on natural language processing and on retrieval-augmented generation in particular. One of those papers measures what a retrieval system does when the documents it is fed have been tampered with.
That is not a footnote. The first AI feature most software firms ship is retrieval over their own documents, and the ways it fails are exactly the ways a demo never fails: the confident answer built on the wrong passage, the citation that points at nothing, the one poisoned document that quietly steers every reply. Knowing that in advance is the difference between a prototype and something you can put in front of a customer.
The rest of the published work runs through machine learning on audio, process mining in clinical settings, and assistive technology. Four fields with nothing in common except the method: take a process nobody has automated yet, find where it actually breaks, build the smallest thing that holds. If your domain appears on none of those lists, that has been the normal case so far, and it is the part of the job we like.
Bring one workflow you want to fix.
A discovery session works best from a concrete process rather than a category. Name the process, name who owns it, and we’ll come back with what it would take to move it.