Selected work
Stage
A mobile app for a Tel Aviv training studio. Every finished workout becomes an Instagram story card with the studio logo, so each trainee post advertises the studio. Built with Claude Code, for a real client.
Case study →
Gate Vision
Shipped for a Bay Area gate installer: specs become a dimensioned, branded drawing plus a photorealistic AI render on the customer’s own house photo.
Case study →
Teach Me This File
An AI tutor that turns a file and a teacher prompt into a Socratic chat, in Hebrew or English. The plan lives in app state, sources come from the plan, scans go to a vision model, and a 16-scenario eval with a judge model checks the tutor.
Case study →
Intent Analysis System
Production pipeline evaluating AI pentesting agents from conversation traces - Sentence-BERT pre-scoring, Claude structured extraction and LLM-as-judge calibration against human ratings.
Case study →
Lightning - Fantasy GM Agent
A LangGraph + Claude agent that reasons over live Yahoo Fantasy data to recommend NBA trades and waiver pickups. Shipped end-to-end with LangSmith evals as a pre-ship quality gate.
Read case study →
Hebrew sentiment classification
LoRA fine-tuned DictaBERT on 43.6k examples, outperforming GPT-4o zero/few-shot and RAG baselines. Plus a GPT-style transformer built from scratch in PyTorch.
Case study →
World Cup 2026 Forecast
Reproduction of a published soccer-ranking paper, extended into a Monte-Carlo forecast of the 2026 World Cup using half-life weighted Poisson models and squad market values.
Case study →
SDR Agent
Outbound agent that reads leads, drafts personalised emails and sends via Gmail API - with retry/backoff, send caps and a dry-run mode.
Case study →- LangGraph
- GPT-4 Turbo
- Gmail API
- Google Sheets
Experience
- Build analytical models and dashboards for SaaS and technology companies, tracking ARR, NRR, churn and LTV across large datasets.
- Turn complex datasets into decision-ready insights for senior stakeholders across multiple transaction diligence engagements.
- Product analytics at an AI-driven cybersecurity startup.
- Built the Intent Analysis System, an internal pipeline that evaluates the company’s AI pentesting agents from conversation traces.
- Four analytics engines: intent distribution, satisfaction, a skill quadrant and a skill-gap recommender.
- An LLM-as-judge step calibrated against human ratings.
- Launched and ran a retail operation generating several hundred thousand NIS monthly.
- Owned concept, pricing, logistics and team management.
- GPA 91.
- Coursework: Deep Learning, NLP, Machine Learning, Time Series Forecasting, Big Data, Network Analysis and Statistics.
- Fine-tuned DictaBERT for Hebrew sentiment on 43,600 examples, beating GPT-4o zero-shot and few-shot baselines.
- Combat soldier in Egoz; Basic Training Commander and Sergeant.
- Completed a 3-month IAF pilot training course.
About
My background combines business analytics with hands-on AI engineering. At EY-Parthenon I work with SaaS metrics such as ARR, churn and LTV; in my own projects I build and evaluate LLM agents. I try to bring both perspectives to product decisions.
- AI & LLMs
- LangGraph, Claude & OpenAI APIs, tool use, RAG, LoRA
- Deep learning
- PyTorch, transformers, attention, CNNs, LSTMs
- Data
- Python, SQL, Pandas, scikit-learn, Tableau, Power BI