Idan Horowitz

GPA 97.7. Thesis: An embedding-space aware calibration loss function, supervised by Prof. Avigdor Gal.
Coursework: Sparse Optimization, ML for Sequential Decision Making, Time Series & Forecasting, Causal Inference, Advanced Econometrics.

GPA 87.3.
Coursework: Machine Learning 1 & 2, Deep Learning, NLP, Foundations & Applications of AI, Database Management, Data Structures & Algorithms, Data Analysis & Visualization, Distributed Data Management, Software Engineering.

• Researched neural-network calibration for my M.Sc. thesis: developed NeSCal, a training loss that smooths predictions across neighbouring examples in a network’s embedding space, with theoretical bounds and benchmarks across 11 image, text and tabular settings (under review at ICLR 2027).
• Co-authored “LLM Agents Display Human Biases but Exhibit Distinct Learning Patterns”: designed matched decision-making experiments for LLM agents and 110 human participants learning from repeated feedback, and found distinct learning patterns despite similar overall performance.
• Taught Python data analysis and machine learning, developed assignments and mentored project teams.
• Built and tested PySpark and SQL exercises on Azure Databricks for data cleaning, ETL and Spark ML in a Big Data course.
• Taught the Internet of Things course and checked homework for Introduction to Causal Analysis.
• Achieved a teaching evaluation rating of 4.5/5.

• Owned the feature pipeline behind store-closure risk models used in investment reports for paying customers, integrating 15–20 messy datasets per chain (APIs, tabular and geospatial) into one clean database.
• Engineered spatial competition features from overlapping store trade areas that improved held-out classification accuracy by 4 percentage points.
• Cleaned unreliable closure labels by cross-checking independent sources with fuzzy address matching and spatial joins; compared XGBoost models and used SHAP to explain feature contributions.
• Cut spatial feature extraction from 8 hours to 2 (75%) through candidate filtering and precomputation.
• Presented findings and limitations to technical and non-technical stakeholders; wrote a 50-page handover and trained my successor.

• Lead, mentor and advise a team of about 8 volunteers, each helping families reach financial independence.
• Recruited and trained 2 cohorts of 20 volunteers.
• Assisted over 20 families in total.

• Cleaned, organised and analysed Paamonim’s organisational data to derive insights for improving its services.

• Helped families take control of their personal finances and become debt-free.
• Accompanied 2 families through their personal-finance journey, each a 10-month process.