Matteo Bonfanti

Applied AI engineer · Piacenza, Italy

I put AI systems in front of the people who actually have to use them.

I build agents, MCP servers and evals — plus the production plumbing nobody demos: streaming, retries, rate limits, cost ceilings.

Fifteen years as a product designer before that, which is the part that matters. I have spent a career watching where real workflows break. Now I build the systems that survive them.

matteo@mattbonfanti.com See what I have shipped

Currently interviewing for Forward Deployed Engineer roles in Europe

The problem I work on

Most AI projects don’t fail at the model.
They fail at the workflow.

The demo works. The pilot works. Then it meets the twelve undocumented steps the team actually performs, and it quietly stops being used — without anyone filing a complaint.

The gap between the demo path and the real workflow Two horizontal tracks. The upper track, the demo, runs from input to output in three clean steps. The lower track, the real workflow, contains the same three steps plus nine undocumented ones branching between them. The vertical distance between the tracks is labelled as the work of deployment. The demo The actual workflow Input Model Output the export nobody automated the approval in a chat thread the spreadsheet on a desktop the exception only Marco knows the field that means two things the manual double-check the work
Every AI project that quietly stopped being used died in the lower track. Finding it is a research problem before it is an engineering one.

Shipped

Systems, not slides.

Each one written the same way: context, constraint, what I built, what happened after. Including the parts that broke.

The method

Fifteen years of this became a protocol.

I wrote it down because I kept doing the same three things in every deployment, in a cardboard plant and in a data room alike. It became a book. It is really a checklist for finding the lower track.

  1. 01

    Deconstruct

    Strip the requirement down to the job being done.

    Most AI projects automate the task that was easiest to describe in a meeting — not the one that was costing anything. The first pass throws away the brief and rebuilds it from what people actually do.

  2. 02

    Decode

    Map the system the humans already run.

    Including the parts that exist only in someone’s head, the workarounds nobody admits to, and the exception that one person handles by hand every Thursday. This is where deployments die, and it is never in the documentation.

  3. 03

    Validate

    Test against real use before the expensive code exists.

    Thirteen heuristics, each with an observable pass or fail criterion and a severity tied to task impact. Not taste, not opinion — a checklist that two people run separately and agree on.

Published as The UX DeCode (Italian). How it applies to AI deployment

What I work with

The honest version.

I write production code. I also know which half of a deployment isn’t a code problem — that is the combination I am selling.

Build

  • Python
  • TypeScript
  • Swift / SwiftUI
  • Node

Model layer

  • Claude API
  • MCP servers
  • Agent skills
  • Tool use / function calling

Keep it honest

  • Eval harnesses
  • Prompt regression tests
  • Cost + latency budgets
  • Failure-mode logging

Ship it

  • Cloudflare Workers
  • Git-based CI
  • Postgres
  • Native macOS / iOS

Background

Founder, designer, engineer — in that order.

I co-founded and ran an AI startup. I have shipped software into a corrugated cardboard plant, an iris-recognition hardware company and a fashion CAD vendor — places where nobody cares about your framework and the workflow has existed for thirty years.

Master’s in AI. Product design at Politecnico di Milano, where I now lecture. I wrote a book about why interfaces fail.

The longer version
Politecnico di Milano
Politecnico di Milano Product design degree, and now lecturer
SDA Bocconi School of Management
SDA Bocconi Social impact entrepreneurship programme

Looking for someone who can get a system working inside somebody else’s organisation?

That is the job I want. Based in Italy, happy to travel, comfortable being the person in the room who does not yet know the domain.

Write to me Download CV