Data Scientist · AI Engineer
Andrei-Ioan
Bleahu
Turning rigorous mathematics into deployable machine learning .
Data Scientist · AI Engineer
Turning rigorous mathematics into deployable machine learning .
Who I am
I began my career distinguishing myself in Romania's National Physics Olympiads, trained as a physicist at Southampton, and then doing research as a Junior Research Fellow at the University of Tartu publishing in peer-reviewed optics journals. Recognising the deep overlap between computational imaging and AI, I pivoted fully into data science by completing an MSc with Distinction at the University of St Andrews, where my dissertation showing that Bayesian Neural Networks can be compressed up to 90% compared to Frequentist methods was awared 18/20. I combine first-principles thinking with hands-on ML engineering to deliver solutions that are both theoretically sound and production-ready.
Career path
2009 – 2013
Romania
Consistently ranked top 10 nationally per year group; twice selected among the top 20 students of all years eligible to represent Romania internationally — placing in the top 0.25% of my generation in Physics.
2014 – 2019
University of Southampton
Integrated Masters in Physics (2:1). Final-year project with the Optoelectronics Research Centre: experimental analysis of Hollow-Core fibres for phase stability in Quantum Repeater Architectures -- awarded first-class honours (71%) for this project, and a strong 2.1 overall for my degree.
2:1 HonoursAug 2022 – Jul 2023
University of Tartu, Physics Dept.
Developed the Lucy-Richardson-Rosen deconvolution algorithm; improved I-COACH and FINCH holographic techniques. First author on two peer-reviewed papers. Presented at SPIE Photonics Europe 2023 (Prague). Co-organised a 48-hr Computational Imaging Hackathon. Secured a French Government bursary to collaborate with the SINGULAR group at the University of Bordeaux.
2 publications · 1st author2023 – 2025
University of St Andrews
Graduated with Distinction (highest classification), GPA 18.1/20. Dissertation: Practical Bayesian Neural Networks that can be compressed up to 90% with no significant accuracy loss which opens the posibility of deplyoment on edge devices. End-to-end ML workflows with PyTorch, PySpark, and Scikit-Learn on real, anonymized datasets. In CS5959 (End-to-End Machinbe Learning) and CS5939 (Data-Driven Systems) I created end-to-end workflow for the full ML cicle, from problem definition and EDA to tuning models and deploying them
Distinction · Dean's List ×2Technical toolkit
Bayesian & Probabilistic ML
Deep Learning
Classical ML & Data
Computer Vision
Programming
Communication