MikhailKulyabin

Two threads run in parallel right now — industrial AI at Siemens and the medical AI company I co-founded — on top of a doctorate, and six years of mechanical engineering before that.

1. 2024 — present

   Postdoctoral Researcher

   Siemens AGErlangen

   FT UX DES, FT UX AXR, FT UX SII

   Generative AI and LLM systems for engineering workflows: multi-agent copilots, retrieval pipelines and knowledge graphs over enterprise data. I do the development and the operations around them — Python services, data collection and storage in Snowflake with dbt models on top, deployment on Azure and AWS, containers, CI and infrastructure as code — plus the user-experience research that decides whether an engineer trusts what comes back.

2. 2024 — present

   Co-founder and CTO

   VisioMed.AI

   I built the technical side of a regulated medical device end to end, from the idea to the working product: the segmentation and classification models that read OCT and fundus images, the copilot that answers over a patient’s history, the platform and the clinical reporting around them, the integration with hospital systems, and the infrastructure it all runs on. In clinics on three continents.

3. 2022 — 2024

   Doctoral Researcher

   Pattern Recognition Lab, FAU Erlangen-NürnbergErlangen

   I ran the AI track of an international retinal imaging project — OCT, fundus and adaptive-optics data, three institutions across Norway, Germany and Russia — and owned it from the data up: the collection and annotation protocols agreed with the clinics, the training pipelines on the university GPU cluster, and the models themselves, CNNs, transformers, GANs and U-Nets for segmentation and classification. Out of it came OCTDL, an open dataset published in Nature Scientific Data and cited over two hundred times, and twenty more papers.

4. 2018 — 2022

   Computer Aided Design Engineer

   Franke GmbHMoscow and Bad Säckingen

5. 2016 — 2018

   Mechanical Engineer

   State Corporation RostecMoscow

1. 2023 — 2025

   Ph.D. in Computer Science

   FAU Erlangen-Nürnberg

   AI for diagnosing retinal diseases: deep learning and computer vision on OCT, fundus and electroretinography data. Supervised by Prof. Dr. Andreas Maier at the Pattern Recognition Lab.

2. 2019 — 2023

   M.Sc. in Computational Engineering

   FAU Erlangen-Nürnberg

   Numerical methods, scientific computing, image processing. Thesis on generative augmentation of medical data.

3. 2016 — 2018

   M.Sc. in Mechanical Engineering

   Bauman Moscow State Technical University

   Continuum mechanics, finite element and CFD simulation, thermodynamics — ANSYS, LS-DYNA and SolidWorks.

4. 2016 — 2018

   B.Sc. in Project Management

   Bauman Moscow State Technical University

   Project management, business planning, financial analysis.

5. 2012 — 2016

   B.Sc. in Mechanical Engineering

   Bauman Moscow State Technical University

   Engineering mechanics, machine design, manufacturing technology.

The doctorate asked one question in four ways: can a machine read the eye well enough to help the person treating it? Two of the four look at pictures of the retina, two at the electrical signal it produces.

1. Scientific Data

   OCTDL, an open OCT dataset

   Retinal deep learning was being trained on a handful of small, closed datasets. So we built one and gave it away: 2,064 B-scans from 821 patients across seven conditions, each graded by medical specialists, with the raw metadata kept so it can be merged with other public sets. It is the second largest open OCT dataset by participants and conditions, and it has since been used to train ophthalmic foundation models in NEJM AI and Cell Reports Medicine.

   Read the paper
2. Sensors

   Reading the electroretinogram as a picture

   An ERG is a one-dimensional waveform, and the diagnostic information sits in when each frequency arrives, not in the shape of the curve. Converting the signal with continuous wavelet transforms turns it into a time-frequency image an ordinary vision model can classify. Stacking three different mother wavelets as colour channels beat every single-wavelet model.

   Read the paper
3. IEEE Access

   Autism in the retinal signal

   The retina is central nervous tissue, so a neurodevelopmental difference should leave a trace in its electrical response. It does. A gated multilayer perceptron reading the light-adapted ERG directly separated autistic from control children — from a test that takes seconds and needs no imaging.

   Read the paper
4. ICPR

   Counting cones in adaptive optics images

   Adaptive optics resolves individual cone photoreceptors, but counting them by hand is the bottleneck. A generalist segmentation model finds and outlines each cone across the mosaic, and Voronoi analysis of the result gives density, spacing and regularity — accuracy holding from the fovea out to two degrees of eccentricity.

   Read the paper

1. 2nd

   Yandex AI Startup Lab

   Second in the DeepTech track with VisioMed.AI. Twelve teams reached the final out of roughly a thousand applications; the prize paid for pilots at Moscow eye clinics.

   2026DeepTech
2. 2nd

   SNOMED CT entity linking challenge

   Second place on DrivenData for a BERT-based linker that maps clinical notes onto SNOMED concepts. Written up in JAMIA.

   2024553 teams
3. Gold

   SETI Breakthrough Listen

   Kaggle gold medal for separating candidate extraterrestrial signals from radio telescope noise.

   2021768 teams
4. 2nd

   HuBMAP: Hacking the Kidney

   Second place on Kaggle for glomeruli segmentation in whole-slide kidney images.

   20211,625 teams

What the work is actually made of.

Spoken

EnglishC1

GermanB2

RussianNative