PyData Amsterdam 2024
2024-09-18..2024-09-20
Amsterdam, Netherlands
[PyData Amsterdam 2016, PyData Amsterdam 2017, PyData Amsterdam 2018, PyData Amsterdam 2019, PyData Amsterdam 2024, PyData Amsterdam 2025]
| title | author | length |
|---|---|---|
| Keynote: Applied NLP in the age of Generative AI | Ines Montani | 43m |
Describes concrete workflows and outcomes: uses spaCy and Prodigy for semi-automated annotation, applies transfer learning to replace LLM prototypes with small… | ||
| Synthetic Data for Localized Solutions | Atieno Ouma | 35m |
Synthetic data drives localized maternal‑health models for Kenya using simulation and GANs and the pipeline includes fidelity utility and privacy checks plus… | ||
| Polars 1.0 and beyond | Ritchie Vink | 37m |
Polars reaches the 1.0 milestone (July 1, 2024) and adds GPU acceleration (August) via collaboration with NVIDIA Rapids, keeping identical semantics and a… | ||
| Open-source Machine Learning on Encrypted Data | Andrei Stoian | 37m |
Concrete ML compiles Python models to run under Fully Homomorphic Encryption so data stays encrypted end to end; it turns arbitrary functions into additions… | ||
| Introduction to Causal Inference using pgmpy | Ankur Ankan | 31m |
Demonstrates the DAG approach with a hands‑on walkthrough on real data using pgmpy, showing causal discovery algorithms (PC, hill‑climb) that produce very… | ||
| dbt-score: a linter for your dbt model metadata | Matthieu Caneill | 28m |
dbt-score is an open source linter that scores dbt model metadata, written in Python and already applied to a repo with 130,000 lines of YAML. It runs locally… | ||
| Dealing With Names and Addresses Around the World | Philip Blair | 37m |
Names and addresses break common schemas: parsing first/middle/last fails for patronymics, single‑name cultures, different orderings and transliterations, and… | ||
| Hunting unicorns with Network analysis | Valerio Ciotti | 36m |
Constructs a monthly time‑varying global startup network from CrunchBase records (1990–2015, ~36,000 people, ~284,000 organizations, 117 countries) by linking… | ||
| Fusing Machine Learning and Mixed Integer Linear Programming | Meent, Tomaszewski | 38m |
Claim: combining ML and MILP can be a superpower or a disaster depending on details. Examples: the US electricity grid optimization (≈$400 billion/year) and… | ||
| SHAP beyond the standard graphics: co-design of ML-models in earth sciences | Hans Korving | 36m |
www.pydata.org
Discover the transformative power of SHAP values in machine learning as we decode complex insights into actionable information for stakeholder… | ||
| Private targeting strategies | Gilian Ponte | 47m |
Claims two implementable methods that satisfy differential privacy and quantifies the tradeoff: Private Causal Neural Networks trained with DP‑SGD to estimate… | ||
| From mocking to rocking your tests with testcontainers | Barend Linders | 37m |
Claim: testcontainers spins real Docker services from pytest so tests hit actual Redis Kafka MinIO emulators rather than pure mocks. Live demo uses a Redis… | ||
| Going Beyond Copilot with AI Agents | Alex Shershebnev | 35m |
Claims that code assistants miss 95% of developer time spent outside the IDE and that ~30% of accepted completions need fixes, then shows a concrete system… | ||
| How Research Teams Can Deliver Higher-Quality Insights Faster | Vasiliy Kaminskiy | 36m |
Claims a nine-step research process that handles hundreds of requests a year as product portfolio grows from 34 products in 2022 to +4 then +5 more, and… | ||
| Build a personalized Commute virtual assistant in Python w/ Hopsworks & LLM Function Calling | Javier | 38m |
Builds a working Python commute assistant that uses LLM Function Calling to decide whether to invoke two concrete functions: a predictor that queries a… | ||
| How Dimensional is a `pandas.DataFrame`, anyway? | James Powell | 34m |
Claim: a pandas.DataFrame is one-dimensional, an index-aligned collection of Series, not a true 2D array. Demonstrations show Series as strictly homogeneous… | ||
| The ML Monitoring Flow for Models Deployed to Production | Wojtek Kuberski | 36m |
Models degrade in production, median ~20% error increase at 365 days and some >80%, so reliance on tests fails; univariate drift tests produce many false… | ||
| Drift Detection on Irregular Time Series with Multiple Non-Uniform Seasonal Patterns | Vitalie Spinu | 36m |
A practical pipeline decomposes irregular business time series with non-uniform seasonalities using an adapted MIST decomposition, segments recent… | ||
| Algorithmic bias is everywhere (especially at Breeze) - what can we do about it? | Thomas Crul | 47m |
A dating app discovers its recommender amplifies existing biases and documents over 220,000 organized dates; engineers trace it to a pipeline combining… | ||
| Building a Data Platform from scratch | Rodel van Rooijen | 37m |
Starts from a Postgres source on Google Cloud and targets terabytes of events; chooses GCP as the cloud, picks BigQuery as the warehouse, uses Spark for heavy… | ||
| From Data Pipelines to a Data Platform: Embracing Monorepo Architecture | Maico Timmerman | 37m |
www.pydata.org
In this talk we will dive into how we architect our on-premise Data Platform on top of a Python monorepo, which allows hundreds of contributors… | ||
| How to measure a city | Françoise Provencher | 40m |
Project measures Montreal against a 2035 goal of 70% living in complete living environments by computing a fine geohash grid of six indicators (shops, transit… | ||
| Uncertainty quantification: How much can you trust your machine learning model? | Mojtaba Farmanbar | 31m |
Claims that conformal prediction gives provable coverage (pick alpha = 0.1 → 90% marginal coverage), that it is model‑agnostic and distribution‑free, and that… | ||
| Je ne regrette rien: Teaching Machine Learning Models Regret Avoidance | Laura Israel | 35m |
Claims that integrating regret-sensitive loss functions into existing models lets them account for downstream components, implemented by mapping game-tree… | ||
| SLM-powered retrieval to scale freelancers matching at Malt | Jouanneau, Palyart | 32m |
Malt builds a neural retriever using SLMs as a frozen 113M backbone with a 7M trainable head, stores embeddings in Qdrant, and serves approximate… | ||
| Terrible tokenizer troubles in large language models | Sander Land | 34m |
Claims that tokenizers leave rare vocabulary entries that break LLMs and cause hallucinations, shows an automated two-step pipeline (efficient candidate… | ||
| Insights from Two Shipmates sailing in the LLM CTF @ SaTML 2024 | Bauer, Fraunholz | 49m |
Team joins the LLM CTF at SaTML and describes concrete wins and failures: defenses use python and LLM filters, attacks combine encodings, jailbreak prompts… | ||
| Keynote: The Art of Language: Mastering Multilingual Challenges in LLMs | Marzieh Fadaee | 44m |
Community builds a multilingual stack: a Discord of >3,000 contributors from 119 countries collects data over eight months, producing the Aya dataset (65… | ||
| LLM Security 101 - An Introduction to AI Red Teaming | Richie Lee | 34m |
Live demo shows an automated red team tool extracting a hidden password from a defended model, using Pirate (Python Risk Identification Tool) to iterate… | ||
| Differential Privacy Made Practical | Rob Romijnders | 35m |
Differential privacy is shown with concrete choices: noise from a Laplace distribution for provable epsilon bounds, Gaussian only under ε,δ relaxations, and… | ||
| Run a benchmark they said. It will be fun they said. | Vincent D Warmerdam | 36m |
www.pydata.org
This is the story of a fun idea that turned into a huge benchmark before it turned into a rabbit hole.
I was trying to figure out reasonable… | ||
| Keynote: Open-source Multimodal AI in the Wild | Merve Noyan | 40m |
Open-source multimodal stacks run end to end on Hugging Face with one-line pipelines and support for 4bit and 8bit quantization so models fit local hardware… | ||
| Lightning Talks + Closing Remarks DAY 1 | 54m | |
Five‑minute lightning set yields concrete wins and failures: converting PDFs to markdown with images using Gemini and GPT‑4 works for structure but returns… | ||
| Jounai.nl: Playing with New Tech to Reinvent the News | Maarten Sukel | 28m |
Jounai is a live AI-driven news site that actually generates up-to-date articles, cites about 50 public trusted sources, and publishes multiple automated… | ||
| Data Science for Social Good: Making Impact in Resource-Constrained Environments | Karami, Kulakova | 28m |
Correlate runs 3–6 month volunteer projects (4–6 people) that use only open source tooling to deliver concrete results: a GIS visualization that changed policy… | ||
| Almost Perfect: A Benchmark on Algorithms for Quasi-Experiments | Raphael Tamaki | 34m |
Compares Difference‑in‑Differences, Synthetic Control, Meta‑Learners and Graphical Causal Models on three open datasets with 200 simulations per dataset… | ||
| RNNs for data-driven multi-touch attribution at Booking.com | Gulmammadova, Mukherjee | 46m |
Builds an attention-based LSTM that predicts conversion from customer journeys, using co-trained embeddings for categorical features and a softmax attention… | ||
| Alice in the Open Source Land | Magdalena Kowalczuk | 32m |
First-hand account gives concrete outcomes: added ddof to std across four occurrences and pushed a PR that passes tests and gets accepted GIFs; work focuses on… | ||
| pydiverse pipedag: A library for data pipeline orchestration | Martin Trautmann | 34m |
A Python library that wires real pandas polars sqlalchemy and ibis code into a DAG and claims automatic cache invalidation so only changed steps rerun, with… | ||
| Time Series forecasting with NumPyro | Juan Orduz | 54m |
Implements custom probabilistic time‑series in NumPyro that actually recover unseen demand using censoring likelihoods (switching forecasts from flat to upward… | ||
| Polishing Python: Preventing Performance Corrosion with Rust | Mike Kraus | 36m |
Compares Python and Rust by porting a pure‑Python payment handler into PyO3 and packaging with Maturin and a Rust crate for parallelism (rayon). Benchmarks run… | ||
| Understanding Polars Expressions when you're used to pandas | Marco Edward Gorelli | 36m |
Polars expressions are presented as a function from DataFrame to Series that enables lazy execution and optimizer-driven reuse (e.g., shared pl.col().shift(1)… | ||
| Keynote: The Art of Language: Mastering Multilingual Challenges in LLMs | Marzieh Fadaee | 44m |
Community builds an open multilingual stack: Aya dataset: 65 languages of manual instruction-style prompts, plus A collection: 500M examples of translations… | ||
| Boosting AI Reliability: Uncertainty Quantification with MAPIE | Cordier, Lacombe | 49m |
Model-agnostic MAPIE produces conformal prediction sets with formal coverage guarantees (eg 95% coverage claimed) by splitting into train/calibration/test and… | ||
| Causal Effect Estimation in Practice: Lessons Learned from E-commerce & Banking | Danial Senejohnny | 39m |
Claims concrete trade-offs: randomized A/B is ideal but often impossible for finance and e‑commerce, so solutions use inverse propensity weighting with… | ||
| Debugging as an experimental science | Sarah Diot-Girard | 35m |
www.pydata.org
If there is only one experience shared by anyone who ever wrote code, it is debugging. Then, why is it so often a frustrating experience… | ||
| How I hacked UMAP and won at a plotting contest | Jeroen Janssens | 27m |
Hacks UMAP source to save intermediate embeddings, runs it on MNIST and draws every epoch, stitches frames with ffmpeg, and wrangles data with Polars… | ||
| nothing matches | ||