Machine learning · Madrid

Matt ffrench

I'm a machine learning engineer. I've spent about ten years building ML systems that run in production, for medical imaging, agriculture, recruitment and marketing, in London, San Francisco, Hong Kong, Sydney and now Madrid.

Outside work I research how well language models know their own limits, and I help run a couple of meetups in Madrid.

Portrait of Matt ffrench

Work

Experience

Mostly applied ML: computer vision, NLP and speech, taken from prototype to something people use.

  1. 2022 – now
    Machine Learning Lead · creative analytics, Madrid

    Leading ML for a platform that measures how ad creative performs.

  2. 2019 – 2022
    Lead Machine Learning Engineer · myInterview, Sydney

    Video-interview screening: NLP on transcripts and models on voice recordings.

  3. 2018 – 2019
    Machine Learning Researcher · Ultromics, Oxford

    Machine vision for cardiovascular diagnosis from echocardiograms.

  4. 2017 – 2018
    Machine Learning Researcher · Chatbot.com.hk, Hong Kong

    NLP for customer-service chatbots.

  5. 2016 – 2017
    Machine Learning Engineer · Mavrx, San Francisco

    Aerial imagery for crop analytics. Joined after an Entrepreneur First fellowship in London.

  6. 2012 – 2016
    MEng, Engineering · University of Cambridge (Christ's College)

    Specialised in information engineering and machine learning. Summer placements at Bristol Robotics Lab (machine vision) and Severn Subsea Technologies.

Research

ConfidenceBench

Accuracy tells you how often a model is right. It doesn't tell you whether it knows when it's wrong.

EIML workshop, ICML 2026

Evaluating confidence calibration in large language models

Models answer 200 hard multiple-choice questions (spatial reasoning, precise maths, exact lookup and offline knowledge) and say how confident they are. We score them with the Brier score, so being confidently wrong costs more than being unsure.

The leaderboard covers 30-odd models and a human baseline, and I update it as new models come out. Written with Daniel Yang, Xinmeng Huang and Sanyam Kapoor.

Using the dataset

The questions aren't public, so they stay out of training data. If you'd like to run models on it for research, email me with a line on who you are and what you want to measure.

Request access

Writing

On Medium

Occasional pieces on AI and on how to work well.

  1. Oct 2025
    When a model finally learned to say "I don't know"

    What the ConfidenceBench results said about GPT-5's calibration.

  2. Jul 2025
    The art of switching off

    Why finishing a task and actually stopping thinking about it matters more than any system.

  3. 2025
    The productivity fuel tank

    Managing energy rather than hours.

  4. Dec 2023
    Don't work like a caveman

    A five-part series on planning, compartmentalising, time pressure and breaks.

All posts on Medium →

Community

Meetups

Two things I started in Madrid and now run with a group of co-hosts.

Madrid AI Builders

Fridays and Mondays

An informal Friday meetup for people working on AI projects, and a smaller Monday session I run with Pablo La Torre where people bring what they're building.

madridaibuilders.com →

Madrilingo

Sundays, Chamberí

An English–Spanish language exchange in a bar on Calle Santa Engracia. Free, and you don't need to sign up for anything beyond Meetup.

madrilingo.com →

Contact

Say hello

Email is the best way to reach me.

mipffc@gmail.com