Skip to main content

Almpion Ratsakou

Lead AI EngineerData and AI platformsLondon

I build the whole path from data to production AI, and make sure it can be trusted: by the people who use it and by the people who have to sign it off. Mathematician by training, PhD Paris-Saclay, five years in production ML.

Founded a company's analytics function Took AI from first experiment to production, on device and in the cloud Hired and mentored AI engineers and data analysts

What I bring

Built in regulated healthcare, where the bar is highest. Everything here travels.

AI that has to leave the lab

A model is not a product. I take AI from first experiment to something people depend on every day, and keep it honest after release: evaluated the way it will be used, watched for drift, retrained when the world moves.

AI where it has to run

When latency, privacy or connectivity rule out the cloud, the model goes onto the device: smaller, faster, and proven to match the original to the number. Phones first; the discipline is the same for any device.

AI that has to be defended

In high-stakes settings, "it works" is not enough. Models that trace back to their data, results anyone can reproduce, and evidence written to the standard regulators, auditors and boards expect.

Data that isn't yet an asset

Most organisations have more data than they can use. I turn scattered systems into one trusted source the business runs on, and the foundation its AI is built on.

Sensitive data that has to stay safe

Lawful, auditable routes from production data into analytics and AI, so an organisation can learn from its data without exposing its customers or itself.

Answers for people, and for agents

Reporting from leadership down, each customer seeing only its own data, and AI measured like the rest of the business. The same trusted data opened, read-only, to AI agents, so they answer from facts.

Leading the work

I have founded a data and AI function from nothing: the platform, the standards, the hires, and the plain-language updates leadership needs. Hands-on throughout, whether leading a team or brought in to start one.

How I work

What stays the same from one company to the next.

01

Proven, not assumed.

If it ships, it behaves as it did in testing: repeatable experiments, data sealed against leakage, evaluation that mirrors real use, parity checked to the number.

02

One source of truth.

Differences live in configuration, not in copies, and core systems change in stages so nothing downstream notices.

03

Fail closed.

With sensitive data the safe path is the default, and every exception is audited.

04

People where a rule shouldn't decide.

Automate the clear cases; route the uncertain ones to a person, by design.

05

Standards a small team can keep.

Decision records, runbooks, tests before deploy, and clear rules for AI coding agents in production code. Enough process to be safe, never enough to slow things down.

06

Plain language upward.

Leadership gets the trade-off, the risk and the recommendation, not the architecture diagram.

CAREER

2022 – 2026TympaHealthHearing health technology, LondonLead AI Engineer
  1. Oct 2024 – 2026

    Lead AI Engineer

    Function lead for AI, data and analytics: the production models, the AI and analytics architecture, serving across device and cloud. The AI evidence for a medical device classification upgrade. Hiring with the CTO; updates to the senior leadership team.

  2. Apr 2024 – Oct 2024

    AI and Data Solutions Engineer

    Founded the analytics function: data platform, warehouse, first dashboards and integrations across the business systems. Most of the design of the governed research data path.

  3. Aug 2023 – Apr 2024

    Senior ML Engineer

    Designed the ML platform with the Head of AI, implemented across the team. Engineering contributor on a grant-funded R&D programme, from which the patent applications came. First mentoring responsibility.

    Joint inventor, Image analysis system for otoscopy images PCT WO2025181495A1 UK GB2638787A published September 2025, pending. Read the filing

  4. Jul 2022 – Aug 2023

    ML Engineer

    Early models, early platform, early operational reporting. Engineering parts of the data handling and model development SOPs.

2021 – 2022KagoolData and AI consultancy for manufacturing, retail and automotive clientsData Scientist (consulting)

Demand forecasting, invoice categorisation, inventory optimisation, customer sentiment analysis and anomaly detection, from proof of concept to production on Azure. MLOps standards for the internal team and technical advice to client teams.

2017 – 2020CEA · Paris-SaclayFrench national research organisationPhD Researcher

Fast simulation and automated defect imaging for thermographic inspection. Four first-author papers.

2013 – 2016FORTH · University of CreteResearch centre, GreeceResearch and Teaching Assistant

Research assistant on an ERC project on imaging in clutter. Teaching assistant for Python, analysis, differential equations and statistics.

EDUCATION

2020

PhD in Electrical Engineering (Signal and Image Processing)

Université Paris-Saclay

2016

MSc in Applied and Computational Mathematics

University of Crete

2014

BSc in Mathematics

University of Crete

RESEARCH

From ultrasound array imaging in clutter to fast simulation and automated defect imaging for thermographic inspection: the signal and image processing foundations behind the computer vision I build today.

THESES

2020
Multi-physical modeling of thermographic inspection methods and fast imaging

PhD thesis, Université Paris-Saclay (CEA LIST)

A fast semi-analytical 3D model of thermographic inspection, and an automated pipeline that detects delaminations in noisy thermograms and estimates their depth and thickness.

More about the PhD thesis
Problem
Finding delaminations between metal layers in noisy thermographic images, and estimating how deep and thick they are.
Approach
A fast semi-analytical 3D model of the inspection, then an automated pipeline: TSR denoising, Canny edge detection and regularised least-squares estimation.
Result
Close agreement with commercial FEM (COMSOL) at a much lower runtime, and under 4% relative error on depth and thickness across the noise levels tested on simulated data.
Today
The same shape as an ML imaging pipeline: denoise, detect, segment, estimate, with physics-based synthetic data to test robustness.
2016
Coherent imaging in clutter

MSc thesis, University of Crete

Ultrasound array imaging through cluttered media: a windowed beamformer equivalent to coherent interferometry and 5 to 10 times faster in the tested setups, and a filter that suppresses multiple scattering.

More about the MSc thesis
Problem
Imaging small reflectors, such as early-stage tumours, with an ultrasound array when their echoes are buried in scattering from a cluttered medium.
Approach
Compared Kirchhoff migration with coherent interferometry, implemented an equivalent windowed beamformer, and added a filter that keeps the single-scattering part of the data.
Result
The windowed beamformer ran 5 to 10 times faster than coherent interferometry in the tested setups, with comparable images, and the filter sharply improved signal-to-noise in strong clutter.
Today
Medical imaging from the start, and the same instincts I use now: make the algorithm cheaper to run, and robust to noise.

PEER-REVIEWED PAPERS

Best Poster Award, ENDE 2017

Outside work

Experiments, none of them public yet: an iOS recipe app with a bundled catalogue and a synced personal library, now gaining on-device language models (Gemma and Apple's foundation models), and a household finance dashboard that gives two people one combined view of their money, built with the testing and security toolchain a large service would have.