WORK, PROTOTYPES AND DEMOS
Some are finished work, others show how we approach a problem. Every card says clearly which kind it is.
The demos don’t collect requests or take payments. The demo websites belong to fictional Italian businesses; the control room demo is available in Italian only.When AI enters a complex business process, handing everything to a single assistant makes it hard to know who did what, with which information, and where to step in before a mistake.
We built an architecture in which specialized agents carry out distinct tasks, pass each other verifiable outputs and ask for human approval at the critical steps.
Procedures, decisions and constraints can be spread across different documents or even contradict each other. An agent needs to know which source prevails and which actions it cannot take.
We built a governed memory for agents: structured documents, a hierarchy of sources, derived indexes and automatic checks make sure information and constraints are applied consistently.
When market information comes from different sources, collecting a lot of it is not enough: you need to know what it actually supports, how complete the research is and where the evidence is weak.
We built a pipeline that gathers different sources, measures how well the research covers the question, keeps the evaluation criteria separate and brings the result to a human decision instead of hiding the uncertainty.
Producing analyses and documents with AI is easy; checking that every claim is actually backed by the sources takes a research and review process you can control.
We built a pipeline that researches and weighs the sources before writing, links claims to evidence and uses separate agents to critique, verify and re-examine the result before it is closed.
Turning technical documents, research or verified material into shorter content can introduce simplifications, claims stronger than the original or information out of context.
We built a system that keeps every piece of content linked to its source, checks how claims are transformed and flags or blocks those that go beyond the defined criteria.
When many pages, product sheets or documents must respect precise sources and rules, manual review gets slow and can let through exactly the claims that matter most.
We built a system that extracts claims in a structured way, links them to their context, compares them with reference sources and measures its performance against independent annotations.
Documents, management software, Excel and email each hold a piece: nobody has the picture of what exists, what is true right now and what is changing, and the status is rebuilt by hand.
We built a system that keeps an up-to-date model of the company and also knows how it is running itself: which sources it reads, what it is allowed to do, what it costs and when to stop and ask a person. Complete and tested, not yet installed at a client of the studio.
The procedure exists, but only one person knows it: everyone else interrupts her or searches through folders.
Three questions, three answers with a source to open and check. On the third, the assistant says it doesn’t know instead of making something up.