I design and build AI agents that plug into the systems your company already runs. And when it's needed first, I help you and your team work out where AI actually pays — and where it doesn't.
Half a day on the process that hurts: who does it, how long it takes, where it breaks. That's what decides whether an agent makes sense — and if it doesn't, I'll say so.
02 — A prototype that runs
I build the smallest version that proves the value, on your real data. You see it work before committing to the rest.
03 — Into production
The agent goes live on your systems, with the autonomy limits you set and a record of what it does.
04 — Kept standing
Things change: processes, vendors, models. An agent has to be watched, corrected and updated over time.
Work
What I've built, and how far it got.
Each card says exactly what stage it's at: in production at a client, built and working, or still in discussion. No generous labels.
Staff scheduling agent
In production
Builds staff rotas around availability, contracts and minimum coverage.
Collects staff availability and time off
Generates a rota proposal that respects contract rules and minimum coverage
Flags gaps and conflicts before publication
Rebuilds the plan when availability changes
Google Cloud EU · Firestore · Cloud SQL
Voice agent for the phone
Built
Answers the phone in Italian, works out what the caller needs and books the appointment.
Answers in Italian, in a natural voice, in real time
Discloses that it is an automated system, as the AI Act requires
Qualifies why the person is calling and captures the useful details
Books the appointment or hands the call over to a person
Gemini Live (Vertex AI, EU) · Telnyx · Google Calendar
Grants and public funding agent
In discussion
Watches grant and incentive calls, keeps only the ones the company is actually eligible for, and warns before deadlines.
Monitors the sources where calls and incentives are published
Discards the ones that don't fit by sector, size or region
Summarises requirements, amounts and deadline in a readable brief
Warns when a window is about to close
Vertex AI (EU) · Cloud Run
Safety compliance deadline agent
In discussion
Tracks medical checks, training and periodic inspections, and warns the right person before they expire.
Connects to the existing system through a dedicated MCP server
Tracks deadlines per person, per asset and per site
Warns the person responsible early enough to act
Leaves an audit trail of what was notified and when
MCP · Vertex AI (EU) · Cloud Run
Work in discussion means a project still being scoped: I describe the problem and how it works, not the client. Client names appear only with their written consent.
Frequently asked questions
The answers you're looking for, in plain terms.
SmartVolve is Mattia Carruggio's practice: it designs and builds custom AI agents for companies, and provides AI consulting and training. It isn't a platform to buy or a subscription product: each agent is built on one specific company's process and systems.
ChatGPT answers whoever types into it. An agent works on its own inside your systems: it reads where the real data lives, does something (books an appointment, extracts data from a document, warns about a deadline) and leaves a record of what it did. In practice: you have to use the first one; the second works when you're not watching.
I don't publish a price list because it would be a number invented before knowing the problem: an agent that reads documents and one that answers the phone are different jobs, and integrating with an open ERP costs far less than with a closed one. Pricing is set after discovery. If budget is a constraint, saying so early saves us both time.
With a call where you describe the work that hurts. If it looks like a good candidate, we look at the process with the people who actually do it, and that tells us whether building something makes sense. If it doesn't, I'll say so: better to find out in a call than after a project.
The prototype comes early by design, because its job is to help you decide: you watch it run on your real data before committing to the rest. Production timing almost always depends on the integrations, not on the AI.
On Google Cloud infrastructure in Europe. The agent accesses what it needs to do its job and nothing else, and every action leaves a trace. Processing is governed by a written agreement: the templates are published in the compliance section.
It will, the way a person does. That's why you start at low autonomy — the agent proposes, someone approves — and raise it only when the numbers say it holds. Processes where a mistake is irreversible are not good candidates to start with.
No. The constraint isn't size but the volume of repetitive work: if something happens twenty times a month and a qualified person does it, it's worth a look — in a company of ten as much as in one of a thousand.
Frequently asked questions about SmartVolve.
Let's talk
Tell me the problem. If an agent isn't the right answer, it's more useful to find out now.
Describe the work that eats your time or keeps breaking. I'll tell you whether it can be automated, what it would take, and how much work it is. If it doesn't make sense, I'll say so and we both save the time.
