AI strategy & agent engineering — Baveno, Italy
I map how your company actually operates, pick the few processes where AI pays for itself, and build the agents that run them — connected to the tools you already use, with a human still in charge.
A real pipeline shape: intelligence → strategy → build → operate. Every step logged, timed and costed.
Why most AI projects stall
Almost every company I meet has already tried something: a chatbot nobody uses, a licence nobody opens, a pilot that never left the meeting room. The technology was rarely the problem.
AI was pointed at a tool instead of at a workflow. Without knowing who does what, when, and with which document, an agent has nothing solid to stand on.
A model that can't read your ERP, your inbox or your price list is a toy. The value is in the integration, and that's the part demos skip.
Agents drift, inputs change, edge cases appear. Without monitoring and a person responsible for corrections, quality quietly decays until people stop trusting it.
What I do
Small, senior, hands-on. You talk to the person who designs the system and to the person who builds it — they're the same person.
I sit inside your operation, map the workflows that consume the most hours, and come back with a ranked list: what to automate, in what order, what it costs, what it returns. No slideware — a plan you can execute or hand to anyone else.
Agents that read documents, answer customers, prepare quotes, chase data, write drafts and update your systems. Built to a defined scope, tested on your real cases, released with guardrails and a human approval step wherever the stakes justify one.
The unglamorous part that makes the rest work: connecting CRM, ERP, mailbox, e-commerce, spreadsheets and internal databases so that data moves on its own and people stop re-typing it. Agents plug into that plumbing, not around it.
For companies that want the capability in-house: I run the roadmap, keep the systems healthy, train your team to use and supervise them, and report monthly on hours saved and cost per run. You end up owning it, not renting it.
Proprietary platform
Anyone can call a model. The hard part is running dozens of agents for months without losing control of what they did, what it cost and whether it was any good. Brainvision OS is the system I build clients on: pipelines of specialised agents, every run recorded end to end.
Illustrative view. Your installation is configured around your own processes and data.
How an engagement runs
The sequence matters more than the tooling: nothing gets built before it's been measured, and nothing scales before one agent has proven itself in production.
Interviews and observation across the teams that carry the load. Every recurring process written down with volume, time and cost.
1–2 weeksA ranked backlog of automations scored on payback and effort, plus the architecture, the data rules and the budget.
1 weekOne agent, one real process, into production. Tested on your historical cases and measured against the manual baseline.
4 weeksIntegrations to your systems, guardrails, documentation and training. Your team learns to supervise, not just to click.
4–8 weeksMonitoring, corrections, monthly reporting on hours saved and cost per run — then the next process on the list.
OngoingWays to work together
Fixed scope, fixed fee, clear deliverables. Figures are agreed after the first call, once we both know what the work actually is.
For companies that want to know where AI pays before spending on it.
A defined system, built and delivered into production on your infrastructure.
Senior AI ownership without hiring a full-time head of AI.
Where this works
Not for AI startups. For manufacturers, brokers, agencies and retailers where a handful of repetitive processes quietly eat the week.
Working in English, Italian and French, remotely across Italy, Switzerland and France — on site for audits and workshops.
Talk about your caseWho you're hiring
I work with companies in Italy, Switzerland and France that are past the curiosity stage and want AI to show up in their operations, not in their marketing. The work usually starts with a process nobody enjoys — quoting, reporting, data entry, first-line support — and ends with an agent doing it every day under supervision.
On the engineering side I design multi-agent systems and build them on Brainvision OS, my own orchestration and observability platform: pipelines of specialised agents, full run traces, cost tracking and human approval gates. That platform exists because I got tired of AI systems nobody could audit six months later.
On the strategy side I stay deliberately conservative. If a process is better fixed with a spreadsheet, a rule or a phone call, I'll tell you — and you'll keep the budget for the one that genuinely needs a model.
Automation is only worth it when someone can still explain, in one sentence, why the machine did what it did.
Questions I get every week
No. Agents take the repetitive parts — the reading, retyping, chasing and first drafts. The people stay for judgement, relationships and exceptions, which is where they were useful in the first place. Every project defines up front which decisions stay human.
An audit takes two to three weeks. The first agent is usually in production within about 30 days of the build starting, on one narrow process. Full rollout across a department typically runs three to six months, one process at a time.
European infrastructure by default, with data-processing terms in the contract. Your documents are not used to train anyone's models. For sensitive workloads we can run open models on dedicated EU infrastructure, and the architecture is agreed with you before anything is built.
Almost never. Agents connect to what you already run — ERP, CRM, mailbox, e-commerce, file storage — through APIs or exports. If a system is genuinely closed, we work around it rather than launching a migration you didn't ask for.
The audit is a fixed fee agreed in advance. Builds are quoted as fixed scope once the blueprint exists, so you approve a number, not an hourly meter. Ongoing support is a monthly retainer. Third-party costs — model usage, hosting, licences — are billed to you directly at cost and never marked up.
Then we stop, and you still own the process map, the measurements and the code produced. The pilot is deliberately small so a negative answer costs weeks, not quarters. That's the whole point of running it before the rollout.
Contact
A first call takes about 30 minutes: your process, an honest opinion on whether AI is the right answer, and — if it is — what a first step would look like. No deck, no obligation.
Cookies and statisticsWe use technical cookies, needed to run the site. With your consent we also use Google Analytics to understand which content is actually useful. You can decline without losing anything. Read the policy