PyData Tel Aviv 2025
| title | author | length |
|---|---|---|
| Fine-Tuning Small Language Models with Python | Sigal Shaked | 28m |
Shows that small language models (0.5–13B params) can match or beat large LLMs for domain tasks using concrete tooling and datasets: trains a 1B model on 120k… | ||
| Admiral-Driven ML Framework for Marine Operations | Teddy Lazebnik | 30m |
Claims an IRL-based metric learning pipeline recovers admiral-level reward functions from demonstrations plus AB-style preference comparisons and sparse… | ||
| Evaluating Your AI Agent | Linoy Cohen, Shirli Di-Castro Shashua | 33m |
A running wine-recommendation agent demonstrates tool calls (Google Maps, retrieval/RAG), dialogue turns and planning failures, and evaluation highlights three… | ||
| Learning How to Learn in the AI Era | Mor Hananovitz, Ortal Ashkenazi | 22m |
Prototype notebook becomes a production-ready recipe agent that actually works: it queries the web, returns at most three recipes, summarizes ingredients and… | ||
| beyond the names - A social trend analysis of name themes | Ira Yaari | 15m |
Uses Israeli Central Bureau of Statistics and a public registry of given names, calls to OpenAI via Python and scraping of common name sites to build a… | ||
| Engineering Production-Ready Onboarding Agents | Noa Radin | 25m |
Describes building a production onboarding agent with an internal Python Agents SDK, Snowflake/Cortex-backed retrieval, and Temporal workers for orchestration… | ||
| Let Your Data Tell Its Story | Ilana Makover | 27m |
Claims a working, low‑overhead in‑house data‑lineage pipeline that uses Airflow Python Redis, sends deduplicated messages from central entry points, parses… | ||
| Putting the "fun" in function calling | Roi Tabach | 14m |
Shows how function calling (passing formal signatures) replaces brittle JSON parsing and saves teams from fragile regex and retry hacks, with examples: a… | ||
| Mining Parliamentary Gold - Building Hebrew ASR | Yanir Marmor, Yoad Snapir | 13m |
Shows a reproducible Python pipeline that remuxes with light ffmpeg, uses stable_ts + faster-whisper + CTranslate2 to force-align noisy Knesset audio and… | ||
| Using DBT and Python to Build Production Pipelines in Snowflake | Ori Cohen | 21m |
A production ML pipeline converts notebooks and legacy scripts into runnable assets using DBT + Python in Snowflake, with models executed inside Snowflake and… | ||
| From Notebook Chaos to Production Gold | Yuval Gorchover | 22m |
A churn model that scores great in a notebook fails in production because training sees full, backfilled tables while serving sees partial, delayed rows and… | ||
| Tabular data Transformed? TabPFN Brings Deep Learning to the Table | Noa Henig | 30m |
Tab-PFN is a transformer-based model for tabular classification and regression that claims dramatic gains by pretrained on millions of synthetic datasets using… | ||
| A New Kind of Learning Systems | Yoav Goldberg | 32m |
Argues that instead of updating model weights, future systems learn by writing, retrieving and editing natural‑language facts in memory, proposing an… | ||
| Is This Feature Interesting? Interfeat: AI for Insight Discovery | Dan Ofer | 27m |
InterFeat is a modular pipeline that filters thousands of tabular features by utility filters (p‑values, mutual information, ML importance), novelty filters… | ||
| Faster Pandas: Speed Up Your Code, Shrink Your Cloud Bill | Miki Tebeka | 31m |
Shows concrete wins: switching CSV to parquet with pyarrow drops an 8.6s CSV load to ~283ms (≈28x); using list appends plus one concat or pandas.from_records… | ||
| Integrating LLMs with Traditional Data Analysis | Maria Murashova | 22m |
Shows concrete integrations and results: LLMs as feature extractors feeding an SQL-backed ML pipeline, synthetic data generation to fix class imbalance, and… | ||
| Talk Less, Graph More: NLP, Networks and Hamilton Musical | Tal N. Mizrachi | 15m |
Builds a song-as-node corpus in Python with networkx, tokenizes lyrics with n‑grams, ranks phrases by TF‑IDF and enforces a minimum of four tokens to reduce… | ||
| A Game-Theoretic Perspective on the Recommender (Eco-)System | Omer Madmon | 26m |
Main claim: concave activation functions guarantee that no-regret dynamics converge to a Nash equilibrium, proven under a formal game-theoretic model where… | ||
| Revealing the Unseen - Leveraging XAI for Deeper Data Insights | Gal Benor | 26m |
Claims a production XAI pipeline that explains fraud scores for >400M (half‑billion) PayPal accounts, grounding explanations in Shapley values and running… | ||
| Recreational Image Reconstruction with Decision Trees | Daniel Anderson | 14m |
Pipeline that maps pixel coordinates to RGB with per-channel decision trees, then replaces XY with distances to implanted sensors using Poisson-disk sensor… | ||
| Tabbers - turn every guitar playing video into a guitar lesson | Lior Kupfer | 16m |
Takes an iPhone guitar performance with noisy, polyphonic audio and distant video and produces time-aligned tablature and an overlayed lesson-ready video… | ||
| Can LLMs do research? | Tal Ifargan, Lukas Hafner | 27m |
Data-to-paper claims an LLM-driven pipeline that converts annotated CSVs into executable Python analysis, figures, and full manuscripts with programmatic… | ||
| How to Build AI Agents and Keep Your Sanity | Shuki Cohen | 27m |
Claims a practical framework that raises agent reliability while containing cost: reflection validation (Pydantic or OpenAI structured outputs and… | ||
| Ummm, Actually | Dean Langsam, Tal N. Mizrachi | 29m |
A live game show quizzing Python data facts, where contestants spot the single wrong detail in claims about Pandas, NumPy, Seaborn, scikit‑learn and Pandas 2.0… | ||
| Marimo - A new notebook | Reuven M. Lerner | 23m |
Marimo claims predictable, spreadsheet‑like single-assignment variables that auto‑percolate changes through dependent cells, builtin widgets (sliders, radios)… | ||
| nothing matches | ||