PyData Boston 2025
2025-12-08..2025-12-10
Cambridge, United States
| title | author | length |
|---|---|---|
| Save your API Keys for someone else | Ian Stokes Rees | 01h 35m |
Demonstrates a working, 100% laptop based set of LLMs using Olama plus Hugging Face model downloads and Llama CPP, pulls ~5–6 GB of models over Wi‑Fi, and runs… | ||
| Build Your MCP server | Chuxin Liu, Yiwen Liu | 01h 16m |
Hands-on workshop builds an MCP server from scratch to expose concrete tools — examples include read unread emails and send email — using the FastMCP SDK and… | ||
| Building LLM Agents Made Simple a | Eric J. Ma | 01h 28m |
A hands‑on build of a back‑office automation system using LlamaBot inside Marimo notebooks maps workflows into three concrete agents: receipt processor that… | ||
| CUDA Python Kernel Authoring | Katrina Riehl | 02h 51m |
Deploy steps use Brev and a GCP L4 GPU, CUDA 12.8 instance, then install the CUDA toolkit (apt repo) because pip alone misses NVRTC. Examples use CuPy and cuDF… | ||
| Generative Programming with Mellea: from Agentic | Nathan Fulton | 01h 24m |
Mellea is an open‑source Python library that replaces giant prompt essays with small imperative calls, explicit instruct–validate–repair steps, and pluggable… | ||
| Notebook to Pipeline: Hands-On Data Engineering w Python | Gilberto Hernandez | 01h 09m |
Claims a complete notebook-to-pipeline run with live code and reusable templates, using Python, Snowflake and Snowpark, plus a 120‑day Snowflake trial for… | ||
| Going multi-modal | Isaac Godfried | 01h 08m |
Hands‑on coverage of model selection, embedding storage, fine‑tuning and production deployment with two concrete builds: a Cairo Geniza search system that… | ||
| Hands-On with LLM-Powered Recommenders: Hybrid Architectures | Sheetal Borar | 57m |
Builds a working movie recommender in Colab that actually retrieves candidates with sentence-transformers and a Faiss index, then sends the top‑K (usually 25)… | ||
| Learn to Unlock Document Intelligence with Open-Source AI | Ming Zhao | 01h 25m |
Docling is an open‑source Python library that actually converts messy PDFs docx excel powerpoint and HTML into a metadata‑rich Docling document format and… | ||
| Understanding & using color for storytelling in data visualization | Lear, Vincent | 01h 16m |
Workshop claims that human perception relies on luminance more than hue so effective color maps use uniform lightness trajectories rather than rainbow ramps… | ||
| The SAT math gap- gender difference or selection bias | Allen Downey | 30m |
A Bayesian model uses PreliZ for priors, PyMC for inference and ArviZ for validation to estimate latent efficacy, a logistic selection curve, and a Gaussian… | ||
| The Lifecycle of a Jupyter Environment | Dawn Michelle Wages | 36m |
A single notebook grows into a runnable ETL pipeline with an actual ETL.py class that registers steps for extract validate clean and execute, backed by tests… | ||
| Where Have All the Metrics Gone | Rebecca Bilbro | 41m |
Explains how RAG systems and LLM agents fail in concrete ways and how to measure each failure: domain failure (authoritative but subtly wrong, only catchable… | ||
| What’s Next in AI for Data and Data Management | Lisa Amini | 45m |
Claims that agents can take a topic description and build, evaluate and iterate a data stack are demonstrated with a demo pipeline that generates schema +… | ||
| Keynote Speaker-Isabel Zimmerman | Isabel Zimmerman | 43m |
Compares pandas to polars (Rust backend) and claims a simple switch speeds common workflows, describes building an MLOps package (Vetiver‑style) and an IDE… | ||
| The Boringly Simple Loop Powering GenAI Apps | Sebastian Wallkötter | 39m |
Claim: most GenAI apps reduce to a boringly simple nested while loop that alternates a persistent turn loop (user turns accumulate history) with an inner trace… | ||
| Embracing Noise- How Data Corruption Can Make Models Smarter | Aayush Gauba | 23m |
Three controlled CIFAR-10 experiments test specific claims: adding Gaussian feature noise during training yields a small robustness gain (clean baseline ~57%… | ||
| No Cloud- No Problem. Local RAG with Embedding Gemma | Sanjit Paliwal | 24m |
Claim: full RAG and topic classification run locally using Embedding Gemma (308M parameters) with Sentence Transformers and Hugging Face, storing vectors in… | ||
| Tracking Policy Evolution Through Clustering | Sarthak Pattnaik | 34m |
Introduces a reproducible pipeline that converts continuous macro series into categorical trajectories using quartile-based categorization, computes similarity… | ||
| Building Production RAG Systems for Health Care Domains | Patel, Doshi | 42m |
Claims concrete: a RAG corpus of more than 7,800 clinical publications and cross‑specialty validation, vector retrieval optimizations that deliver up to 35%… | ||
| LLMOps in Practice- Building Secure, Governed Pipelines 4 LLMs | Siddharth Shankar | 36m |
Claims a reproducible three‑tier LLMOps pipeline (sandbox → runner → production) with a working demo app that shows RBAC, strict prompt templates, logging and… | ||
| Evaluating AI Agents in production with Python | Susan Shu Chang | 57m |
Uploader description shows a Python-based evaluation framework that runs in production for over a year and actually detects regressions. It uses tracing with… | ||
| Fun With Python and Emoji | Ted Conway | 43m |
Proposes using emoji as syntax and claims a workable pipeline: VS Code + Jupyter extension uses an IPython input transformer plus a regex emoji detector to map… | ||
| Applying Foundational Models for Time Series Anomaly Detection | Abhishek Murthy | 41m |
Compares two concrete pipelines: TimeGPT (forecasting, uses confidence intervals from a model trained on ~100 billion points, flags observations outside the… | ||
| Three agents, three frameworks, one talk | Benjamin Batorsky | 41m |
Compares three concrete implementations for agent‑based search and shows what actually works: a hand‑rolled agent that uses Llama 3.1 (7B) locally and exposes… | ||
| Patterns for Productive Agent-Assisted Programming | Eric J. Ma | 52m |
Claims a repeatable four-step workflow—plan your changes, write tests first, let the agent build, document what you created—and shows it on a real repo using… | ||
| Scaling Specialist Knowledge with AI | Fasal Shah | 38m |
Red Hat builds a multi-tool AI agent that moves beyond static RAG to call tools via a Model Context Protocol, pulling live account signals from People.ai… | ||
| Using Cursor (and other AI code gen tools) for data science | Mike Woodward | 41m |
Claims that Cursor and similar AI IDEs produce production‑quality data‑science code while also failing on important tasks. Demos use Cursor, Claude, Bokeh and… | ||
| How AI Is Transforming Data Careers | 46m | |
Panelists report concrete shifts: roles move from feature engineering to building AI products, writing prompts and orchestrating workflows using GenAI and LLMs… | ||
| Processing large JSON files without running out of memory | Itamar Turner Trauring | 37m |
Loading a 50 MB JSON file can use ~230 MB of RAM (about 4.5×) and a 10 GB file can demand ~40 GB, so peak memory is the limiter not average. Four practical… | ||
| Accelerating Geospatial Analysis with GPUs | Jaya Venkatesh | 36m |
An end-to-end pipeline loads Sentinel‑2 tiles and ESA WorldCover labels, preprocesses clouds and overlaps into a median 94.34GB → ~200MB, 240 chunks Zarr… | ||
| Wrappers/Extenders- Companion Packages for Python Projects | Jules Walzer-Goldfeld | 32m |
gt-extras extends Great Tables by shipping 3–5 concrete utilities: dumbbell and bar mini-plots rendered as SVGPY-generated SVGs, a one-line highlight-column… | ||
| MMM Open- Source Showdown- A Benchmark of PyMC-Marketing vs. Google Meridian | Luca | 40m |
An open benchmark compares PyMC‑Marketing and Google Meridian on ESS/s, parameter recovery, convergence diagnostics and memory use using synthetic multi‑geo… | ||
| Modeling Aesthetic Identity- Building a Digital Twin | Pranav Kompally | 42m |
A reproducible pipeline called kite architecture converts Instagram likes, followed brands and ignored posts into OpenCLIP visual embeddings plus categorical… | ||
| When Rivers Speak Analyzing Massive Water Quality Datasets | Rodrigo Silva Ferreira | 39m |
Claims that terabytes of sensor records become tractable by combining USGS Water Data APIs with Remote SSH in Positron and columnar storage. The workflow… | ||
| Rethinking Feature Importance-Evaluating SHAP & TreeSHAP for Tree-Based MLM | Gao | 34m |
Compares SHAP’s assumptions and practical limits with concrete experiments on XGBoost LightGBM CatBoost models, showing kernel SHAP’s sensitivity to label… | ||
| Unlocking Smarter Typeahead Search | Brandon (Anbang) Wu | 36m |
Shows a reproducible hybrid: a low‑latency prefix table plus semantic retrieval built with open source tools and deployed on GPU, encoding 200 million titles… | ||
| Measuring Media Impact- Practical Geo-Lift Incrementality Testing | Bryce Casavant | 36m |
Explains how running geo-lift incrementality tests on DMAs or similar geos gives a privacy-agnostic causal estimate, using an AA check, market stratification… | ||
| fastplotlib-scientific discovery through data visualization | Lewis, Kolar | 33m |
Claim: a GPU‑backed plotting library called fastplotlib delivers very fast, nonblocking scientific plots by building on pygfx via WGPU and modern APIs like… | ||
| One agent, one job, better A | David Jones-Gilardi | 37m |
Claim: breaking tasks into narrowly defined agents produces more accurate, consistent, and cheaper results. Builds an Langflow app wired to LM Studio models… | ||
| Who is Python for- EVERYONE (and why that matters) | Deb Nicholson | 39m |
Python runs on a nonprofit model that keeps the language and its package index free and vendor neutral, with governance open to community proposals and monthly… | ||
| The Column's the limit | Paddy Mullen | 43m |
Buckaroo is a concrete, open‑source notebook table that actually reads big files without filling RAM: it uses viewport streaming, lazy Polars pipelines… | ||
| Surviving the Agentic Hype with Small Language Models | Serhii Sokolenko | 40m |
Claims that compact SLMs can do planning, tool calling, code glue and instruction following, with concrete demos: XLM2 32B quantized runs on a laptop (<16 GB… | ||
| The JupyterLab Extension Ecosystem- Trends from PyPI +GitHub | Konstantin Taletskiy | 39m |
Analysis of PyPI and GitHub data shows concrete signals not vague promises: downloads, repo stars, release recency and categories are measured via BigQuery… | ||
| Is Your LLM Evaluation Missing the Point | Daina Bouquin | 42m |
A concrete mismatch emerges: evaluation suites report 93% accuracy while deployed systems produce 48,000 flagged cases that a review finds 93% wrong and tools… | ||
| nothing matches | ||