Skip to content
FV Francesco Vigni
Francesco Vigni
IT

Hi, I'm Francesco 👋 · High-Tech Artisan

I find out when the model is wrong.

Applied ML Engineer · AI Consultant

PhD engineer and applied AI researcher. I take difficult models from research to systems that run: medical-imaging foundation models, robots on factory floors, and the infrastructure that keeps both in production. Whether it starts as a research question or a production problem, the work is the same.

Let's talk My work
FV Francesco Vigni

Ten years, research & industry

  • Univ. Naples Federico II
  • Univ. Siena
  • TU Wien
  • TU Munich
  • Disney Research
  • Roboception
  • Marie SkĹ‚odowska-Curie Fellow

What I do

Three areas, one throughline: getting AI to production.

Applied AI / ML

Foundation models, computer vision, self-supervised learning, and the model evaluation that says whether a result is real. Medical imaging is where most of it runs.

Robotics & Edge AI

ROS2 navigation and perception, real-time inference on NVIDIA Jetson, embedded deployment, and industrial systems sold in several countries.

ML Infrastructure

Training infrastructure and MLOps: distributed runs, experiment tracking, reproducibility, and Kubernetes delivery that rolls back on its own.

Selected work

Evidence, not slideware.

AI / ML Research

Foundation Model for Gastroenterology Imaging

SSL pretraining strategy and cloud data pipeline for a ViT foundation model on 5M+ gastrointestinal video frames.

5M+ frames

gastrointestinal video, self-supervised pretraining

Read the case study

Robotics & Edge

Edge AI Occupancy Monitoring System

Real-time people tracking and zone-based automation running on NVIDIA Jetson at the edge.

Real-time on Jetson

on-device tracking and distance estimation, MQTT zone events

Read the case study

Robotics & Edge

Reliability Uplift in Industrial Bin-Picking

Core perception and grasping software for Roboception's rc_visard and rc_cube industrial products.

+9% grasp reliability

industrial bin-picking, shipped in rc_visard and rc_cube

Read the case study

Research → All work →

Research

I don't just train models. I test what they're actually learning.

Two recent studies, both on public data, both with the negative results left in.

All research

  • Endoscopy · acquisition shortcuts

    0.961

    as acquired

    You can tell which hospital a colonoscopy frame came from 96% of the time, without looking at the anatomy, and the standard fix barely helps.

    Four findings, and the one that failed

  • Fetal ultrasound · orientation

    0.28°

    second-order moments, no training

    Estimating fetal cardiac orientation does not need a trained model: closed-form geometry beats the network by two orders of magnitude.

    Why landmarks, and where they break

My story

From a robotics thesis at Disney to AI in the clinic.

I studied engineering in Siena, and did my Master's thesis at Disney Research Zurich, on closed-loop control for a robot that shakes your hand, which ended up as a paper in IEEE RA-L. That set the tone for everything since: I like building things that have to work with real people, in the real world.

I spent years in Germany making that happen: autonomous mobile robots deployed across factories, and core 3D-perception and grasping software at Roboception for industrial pick-and-place. Then I went back to research for a Marie Skłodowska-Curie PhD in Naples, with secondments at TU Wien and studying how robots make their intentions legible to people.

Let's talk

Tell me what you are trying to build, what is not working, and what constraints you are dealing with.