Maxim Popov

Machine Learning Engineer

PNHZ (Pavlodar Oil Refinery)

Pavlodar Oil Refinery is a major oil refining facility in Kazakhstan, processing crude oil into fuels and other petroleum products.

Built ML systems for PPE detection, automated annotation and camera tracking, including a pipeline that reduced manual annotation work by 90%.

• Designed a VLM-based agentic pipeline for structured image captioning using ADK and PyDantic in a PPE detection platform. • Deployed Gemma 4 26B on a GPU cluster using a Docker-containerised vLLM; achieved a 4-fold increase in memory throughput via KV cache quantisation and per-conversation token budgeting. • Implemented a custom serverless execution engine to automate image annotation with CVAT using the Gemma agent as a service, reducing manual labour by 90%.

PhD in Clinical NLP

Amsterdam UMC

Amsterdam UMC is a university medical center in the Netherlands that combines patient care, medical research and health professions education.

Develops clinical NLP tools, including Dutch speech transcription and synthetic dialogue generation, and researches alignment of LLM outputs with clinical guidelines.

• Developed a long-form Dutch speaker-attributed ASR pipeline using Qwen Omni for transcription, diarization, and timestamp prediction. • Engineered a pipeline to generate high-fidelity synthetic clinical dialogues to train models while maintaining strict privacy standards. Utilised voice-to-text synthesis and anonymisation techniques to expand training datasets without compromising patient confidentiality. • Researched and applied Group Relative Policy Optimisation (GRPO) and PPO frameworks to align LLM outputs with clinical safety guidelines and structured reporting requirements.

Research Assistant

Nazarbayev University

Nazarbayev University is a research university in Astana, Kazakhstan, with programs across science, engineering, medicine and other fields.

Coordinated clinical data annotation for a 562-patient CT dataset and ran 3D U-Net experiments for medical image segmentation.

• Coordinated clinicians in developing MAGIC-CT, a dataset of 562 patients combining CT scans, 3D organ and lesion annotations, and radiologist-authored clinical descriptions. • Developed annotation guidelines and trained clinicians to use annotation software, supporting consistent image and text labeling. • Conducted 3D U-Net experiments for volumetric medical image segmentation. • Collaborated with researchers at Johns Hopkins University to expand the dataset’s annotations.

Master’s Thesis Student (PI: Dr. Anja Hennemuth)

Deutsches Herzzentrum der Charité (DHZC ICM)

Deutsches Herzzentrum der Charité is a specialist heart center in Berlin, providing cardiac care and supporting cardiovascular research.

Built and evaluated vision-language approaches for coronary artery disease analysis, including diseased vessel classification with an AUROC of 0.93.

• Implemented a novel pipeline for efficient coronary artery disease (CAD) analysis using X-ray coronary angiography images based on vessel centreline extraction. • Proposed and developed a VLM-based approach for the classification of diseased vessels with an AUROC of 0.93. • Assessed the capabilities of general-purpose LLMs and VLMs for the diagnosis of CAD without fine-tuning. • Created a curriculum learning setup by combining and augmenting several open-source datasets. • Performed RL (GRPO) fine-tuning of the PaliGemma 2 model for smart visual analysis of the coronary angiography data.

Computer Vision Researcher

CMC Technologies

Developed a patented pipeline to assess coronary vessel thickness changes and stenosis severity.

• Implemented an automated vessel assessment pipeline, allowing for tracking changes in vessel thickness to detect and assess stenosis severity in coronary vessels (Patent).

PhD, Clinical Informatics

Amsterdam UMC

Master of Science, Computer Vision

Mohamed Bin Zayed University of Artificial Intelligence

Bachelor of Science, Computer Science

Nazarbayev University