Data Scientist | MSc in Data Science from BGU

I'm a Data Scientist passionate about building intelligent systems with Machine Learning and Generative AI. My background combines academic research, data science, and product-oriented thinking. I enjoy turning complex ideas into practical AI solutions and I'm currently exploring LLMs, agentic AI, RAG, and evaluation frameworks.

A simulation framework for collaborative document writing using LLM-based agents, presented at ECAI 2025.

• Problem: Explored how LLM-based agents can collaborate on shared document writing through dynamic propose-and-vote interactions.
• Architecture: Designed and implemented multi-agent simulations in Python, including agent roles, interaction workflows, and collaborative decision-making mechanisms.
• AI Workflow: Built end-to-end Generative AI workflows using OpenAI APIs, LangChain, and RAG, including prompt engineering, agent orchestration, and evaluation pipelines.
• Evaluation: Applied NLP, recommendation algorithms, and data analysis techniques to analyze large-scale simulation outputs, identify behavioral patterns, and evaluate system performance.

An LLM pipeline that classifies B2B support tickets and drafts grounded first-responses, using retrieval over historical precedent.

• Problem: Automated the first pass of support triage - categorizing and prioritizing incoming tickets and drafting a customer response grounded in how similar issues were resolved before.
• Architecture: Built a modular 8-stage pipeline (load → preprocess → classify → validate → post-process → evaluate → error-analysis) driven by a single re-runnable entry point.
• AI Workflow: Combined BM25 retrieval over a deduplicated knowledge base with structured-output (JSON-schema) LLM classification, plus deterministic groundedness flags and rule-based heuristics for reliability.
• Evaluation: Designed an LLM-as-judge harness scoring response quality alongside category/priority accuracy, with per-category error analysis driving iteration.

Enhancing HuBERT-based Speech Emotion Recognition through data augmentation, fine-tuning, and cross-dataset evaluation.

• Problem: Investigated how data augmentation improves the robustness and generalization of Speech Emotion Recognition (SER) models under diverse acoustic conditions.
• Deep Learning: Fine-tuned a pre-trained HuBERT model for seven emotion classes and evaluated SpecAugment, Time Stretch, Additive Noise, and Neutral Copy-Paste augmentation strategies.
• Evaluation: Designed a comprehensive experimental framework comparing individual and weighted augmentation combinations, including cross-dataset validation on the SAVEE corpus.
• Key Findings: Demonstrated that frequency-domain augmentations significantly improved recognition performance while highlighting the remaining challenges of cross-corpus generalization.

An end-to-end Business Intelligence platform combining data warehousing, interactive dashboards, and predictive analytics for railway operations.

• Architecture: Designed a three-layer BI architecture consisting of a Star Schema data warehouse, Tableau dashboards, and Python-based analytics.
• Data Engineering: Built a dimensional data model integrating operational, ticketing, scheduling, and station data into a unified analytical platform.
• Analytics: Applied K-Means clustering and Gradient Boosting to segment station profiles and forecast passenger demand for operational planning.
• Business Intelligence: Developed executive dashboards and OLAP reports to monitor KPIs, operational performance, revenue trends, and passenger flow.

A Dynamic Approach to Collaborative Document Writing   |  

Lead author • Presented at ECAI 2025, an A-ranked international AI conference.