I build machine learning systems end to end, from raw data and preprocessing through modeling and optimization to pipelines that run in production. My work spans computer vision, time series and signals, video, and agentic systems.
My name is Itamar. I'm a Data Scientist with a background in research and software engineering. In recent years I've focused on building Machine Learning and Deep Learning solutions end to end: from raw data and preprocessing, through feature engineering, modeling, and hyperparameter optimization, to scalable pipelines and production systems. I've worked with both classical machine learning and deep learning across a wide range of data types, including structured data, time series, signals, images, and video, primarily using Python, PyTorch, PyTorch Lightning, and scikit-learn.
Before moving into Data Science, I spent several years as a software developer at the Prime Minister's Office, building internal systems, ETL pipelines, and infrastructure for a large data cluster, working with Python, C#, SQL, and a range of monitoring tools.
Alongside that, I built up substantial research experience. During my M.Sc. I led research on analyzing cellphone signals to predict chronic pain and suicide risk. The work involved reading and implementing academic papers, developing new models, and scientific writing, and led to publications at ICMLA, HICSS, and CHIL.
Today I work as a Data Scientist at Teva Pharmaceuticals. I've worked on a range of projects in healthcare and drug development. The main ones were computer vision systems for analyzing biological images, real-time video monitoring systems for assessing patient severity, and LLM-based and agentic solutions for automating processes and generating reports.
I'm looking to keep growing in a strong technological environment, working on complex problems with real impact, and to go deeper into AI, Machine Learning, and production systems. It matters to me to build solutions that don't stop at the POC or research stage, but ones that can actually be used, effectively and efficiently, with sensible use of resources and a proper fit for production environments.
Eight projects across industry, research, and personal AI engineering.
Built and deployed a computer vision system that scores cell-line candidates from microscopy images, helping scientists select high producers for drug development.
Production deployment · ROC-AUC 0.734 Automated scoring to support cell-line selection
Developed a clinical proof of concept that assesses involuntary movement severity from patient videos, translating visual movement into a quantitative score aligned with clinician assessments.
ρ = 0.80 agreement with the in-site clinician A human remote rater watching the same video reaches 0.57
Built an agent that carries clinical trial analysis from raw data to modeling and statistical conclusions, accelerating the workflow while keeping the data scientist in control of key decisions.
Turns about a month of analysis into several days A coworker with clinical knowledge that helps design experiments and validate analyses.
Automated the creation of pharmaceutical validation reports from laboratory data and regulatory protocols, bringing calculations, results, and narrative into a complete report.
Days or weeks per report, down to minutes Orchestrates report creation, from extracting data to composing the final report.
Developed methods for predicting suicide risk and chronic-pain health states from passive cellphone data despite scarce clinical labels, contributing the methodology to two published studies.
Predicting health risks from cellphone usage Introduced a new classical machine learning method for time-series classification, applied to suicide-risk and chronic-pain prediction with limited clinical labels.
Developed a probabilistic time-series classifier that outperformed GP-VAE on Healing MNIST and achieved the best average rank among deep models in the UEA evaluation.
State-of-the-art time-series classification model Achieved the best average rank among evaluated deep models on UEA and 88.2% accuracy on Healing MNIST, compared with 77.9% for GP-VAE.
Extended GP-VIB to learn from both labeled and unlabeled time series, improving classification when labels are scarce and observations are missing.
Outperforms GP-VIB when labeled data is limited Uses both labeled and unlabeled time series to improve classification accuracy, with larger gains when labels are scarce and observations are missing.
Built a private family coordination system that brings messages, tasks, files, and calendar events into shared plans and daily routines.
A household system that automatically extracts tasks from incoming information Turns incoming messages and documents into tasks, shared plans, and daily routines.