PyData London 2026
2026-06-05..2026-06-07
London, United Kingdom
[PyData London 2016, PyData London 2017, PyData London 2018, PyData London 2019, PyData London 2022, PyData London 2024, PyData London 2025, PyData London 2026]
| title | author | length |
|---|---|---|
| Hands-On with Tabular Foundation Models | Nicolas Makaroff | 01h 28m |
Hands-on notebooks run TabICL-style models and compare them rigorously to XGBoost and Random Forest using sklearn, tabicl, xgboost, pandas and an API option… | ||
| Flexible Statistical Modeling | Chris Fonnesbeck | 01h 27m |
Claims that BART gives calibrated probability intervals and that pymc-bart makes it practical get tested with concrete examples: F1 lap times and GSS… | ||
| Making Databases LLM-Ready Building Semantic Layers with Semantido | Dragos Crintea | 01h 28m |
Claims a production-grade semantic layer called Semantido that ties SQLAlchemy models to operational metadata, keeps semantic definitions in sync with… | ||
| Pydantic Monty & Logfire | Samuel Colvin | 42m |
A minimal Python interpreter written in Rust called Monty runs AI-written code with zero builtin syscalls and requires explicit capability grants to call host… | ||
| Do you know how well your model is doing? Evaluate your LLMs | Cheuk Ting Ho | 27m |
Hands-on workshop that shows how to fine-tune a small GPT2 locally on GSM‑8K, run comparisons between two models (GPT2 and Qen) with Lighteval, and plug custom… | ||
| Practical Multicalibration with MCGrad | Niek Tax | 01h 23m |
Claims a practical post‑processing that enforces subgroup calibration by fitting gradient‑boosted additive corrections to model scores using decision‑tree… | ||
| Teach an LLM to Navigate the Web | Richard Kehinde Ogunyale, Oreolorun Olu-Ipinlaye | 01h 10m |
Claims a working browser agent built from scratch using Python Playwright and an LLM API and shows it actually navigates Amazon, adds a book to cart, and posts… | ||
| From Synthetic Examples to Production Signals | Mulepati, Ramaswamy | 01h 30m |
Claims a usable pipeline that turns synthetic multimodal examples into production training signals using NeMo Data Designer and NeMo Anonymizer, with models… | ||
| Don’t Call It 'The Forecast': Structured Prediction at Scale | Thomas Ogden | 21m |
Forecasting at Spotify becomes a structured prediction system: forecasts are versioned objects in a high‑dimensional space with lineage, schema contracts and… | ||
| Test-Driven Data Analysis | Nick Radcliffe | 01h 31m |
Methodology and an open-source Python package tdda deliver concrete results: automatic constraint generation and discover/verify/detect workflows validate data… | ||
| Learn to Unlock Document Intelligence with Open-Source AI | Ming Zhao | 01h 19m |
Docling is an MIT‑licensed Python library that actually converts 15+ file formats into a structured intermediate Docling document using OCR, layout models… | ||
| Model criticism with ArviZ | Oriol Abril-Pla | 01h 19m |
ArviZ demonstrates concrete model-criticism tools: posterior predictive sampling (about 2,000 posterior predictive realizations in examples), marginal overlays… | ||
| GPU Algorithm Authoring with CUDA Tile | Katrina Riehl | 01h 26m |
Claim: tile programming with CUDA Tile and cuTile Python lets authors write array‑centric GPU kernels that stay portable across NVIDIA architectures. The… | ||
| Using coding agents with open models | Sujee Maniyam | 35m |
Claims that open models now sit close to proprietary ones on benchmarks while costing far less and that a hosted service (Token Factory) adds aggressive… | ||
| Build Reasonable Software | Jeremiah Lowin | 33m |
Concrete claims: FastMCP v3 redesign compresses a 2,500‑line PR to 400 lines by reusing abstractions and replaces brittle feature bloat so it actually works… | ||
| Production-Ready AI Agents: From LLMs to Small Language Models | Prattyush Mangal | 36m |
Roadmap claims a practical migration from frontier LLMs to small language models with concrete results: a hybrid Granite SLM achieves 50% footprint reduction… | ||
| Building a Scientific Taxonomy at Scale | Daniele Raimondi, Feichi Lu | 37m |
Pipeline builds a 115K-concept layer attached to OpenAlex and claims per-field graph clustering at scale using SPECTER2 embeddings, mutual kNN graphs, and… | ||
| Evaluating multi-turn conversations | Lena Shakurova | 32m |
Claims a concrete, three‑stage framework — experimentation, live monitoring, post‑production — with implemented Python pipelines, RAG precision/recall and… | ||
| Designing for high-performance execution with Arrow and Polars | Kamlesh Shah | 39m |
Claims concrete speedups: Arrow’s zero-copy Arrow format and Polars’ lazy execution cut serialization and data-movement costs and often give order-of-magnitude… | ||
| MCP, or not MCP | Neal Richardson | 35m |
Claims that MCP is an HTTP API spec that safely exposes narrowly scoped tools to LLMs for real workflows and that fastMCP plus Claude Code can produce a usable… | ||
| Kafka Streaming, the Pythonic Way | Arthur Andres | 36m |
Concrete claim: Kafka is a source of micro‑batches that can be deserialized into Arrow-backed DataFrames so the same DataFrame code runs against Parquet or… | ||
| Documenting your open source projects for machines | Jacob Tomlinson | 37m |
Claims that LLM-driven agents read and prefer machine-friendly docs, with concrete steps and tools: publish markdown in wheels, LLMs.ext sitemap pages, use… | ||
| Designing Semantic Memory for Multi-Agent Systems with Python | Theo van Kraay | 46m |
Argues that semantic memory is a data engineering problem and shows design choices that actually change results: layered memories—short-term conversational… | ||
| Bridging Pandas and Polars | Ivo Dilov | 34m |
Benchmarks show conversion cost is concrete not theoretical: numeric round trips cost about 250–300 ms conversion and Parquet reads into Pandas take ~3 s while… | ||
| JupyterLite: run all your code in a web browser using WebAssembly | Ian Thomas | 39m |
JupyterLite runs entirely in the web browser using WebAssembly, demonstrating CPython via pyodide and xeus-python kernels, compiled C++ notebooks (a ray… | ||
| Reading the Mind of an LLM | Luca Baggi | 35m |
Claims that mechanistic interpretability can locate and manipulate feature directions inside real LLMs — using sparse dictionary learning, circuit tracing and… | ||
| AI sovereignty, provenance and compliance | Daina Bouquin | 33m |
Defines AI sovereignty as reclaiming control and giving concrete steps: cryptographically signed models, dataset provenance receipts and SBOM/AI BOM… | ||
| Zero Trust with Iceberg REST Catalog and Policy Engines | Viktor Kessler | 33m |
Claims a reproducible governance layer: the Iceberg REST Catalog becomes a centralized enforcement point, issuing short‑lived credentials and decoupling access… | ||
| Did Your Rollout Actually Work? | Benjamin Vincent | 37m |
Demonstrates that naive pre/post and standard two‑way fixed‑effects give biased answers for staggered rollouts and that an imputation / staggered DiD approach… | ||
| Hazards on the Causal Path | Nathaniel Forde | 21m |
A Bayesian implementation of Dynamic Path Analysis in PyMC models discretised time hazards with smooth spline priors and a Poisson discrete‑time likelihood… | ||
| Beyond Spark MLlib: Deduplicating Common Crawl at Scale | Ken Obata | 35m |
Presents a production pipeline that deduplicates Common Crawl at scale using partition-aware MinHash LSH that drives Spark partitioning, vectorized MinHash… | ||
| Do Multilingual Embeddings Really Share a Semantic Space? | Kavit Tolia | 23m |
Finds that pre-trained multilingual embeddings mostly align but unevenly: a tiny retrieval demo sometimes works (dog aligns) and sometimes fails (fire emoji… | ||
| Mapping the local heat transition | Sofia Pinto, Simran Dave | 33m |
A reproducible pipeline maps the most suitable low‑carbon heating per property and agglomerates them into neighbourhood Local Heat Planning Tool opportunity… | ||
| From Noisy Sensors to Events | Ono Gantsog | 29m |
A production Python pipeline smooths noisy weight sensor streams with a Kalman filter (filterpy) then infers discrete phases with a Hidden Markov Model… | ||
| Fast-Forward(ing) Models with AI Emulators | Austen Wallis | 37m |
Emulator pipeline replaces costly forward runs of the Sorocco radiative transfer code with a neural-network surrogate that shrinks a 30+ parameter space to… | ||
| Building Production Multi-Agent RAG Systems on Serverless AWS | Samuel Jaja | 23m |
Claims deployment of a production multi-agent RAG platform on serverless AWS with a five-agent financial analysis system called Prime and concrete lessons… | ||
| SELECT instance FROM cloud WHERE workload = ? ORDER BY cost | Gergely Daroczi | 34m |
Spare Cores Navigator is a Python‑queryable benchmark dataset covering 5,000+ instance types and >500 workloads with standardized performance and… | ||
| Decoding and Shifting Team Culture From Any Seat | Margaritha Groenendijk | 44m |
Framework for reading unwritten rules and a clear playbook for changing them from any role claims the talk, with three influence moves you can model… | ||
| LLMs and AI agents demystified | Martin O'Reilly | 36m |
Models primarily predict the next token using embeddings and transformer attention, scaled to hundreds of billions of parameters and trained on hundreds of… | ||
| Vibe NLP for Applied NLP | Ines Montani | 27m |
Claims a concrete workflow: use LLMs to write reusable scripts and training data so runtime inference uses small private classifiers that are faster and… | ||
| The Polars vs SQL differences nobody is talking about | Marco Edward Gorelli | 20m |
Polars treats data as named, same-length columns so row order preserved unless explicitly changed, allows column-wise operations that can produce… | ||
| Querying the queries: SQL Metaprogramming in Python | Michel Semaan | 35m |
Python parses SQL into ASTs represented as nested dictionaries (using libraries like sqloxide) so queries become editable data rather than fragile text… | ||
| The Human-in-the-Loop is Tired | Laura Summers | 34m |
Thesis: optimise for human experience not model output — three named failure modes are reward function disruption, intensity trap, and isolation drift… | ||
| The Future of Notebooks in a Claude Code World | Paddy Mullen | 27m |
Claims a working Deconstructed Notebook that replaces kernel state with automatic caching to disk, using the PyData Arrow stack (Ibis, Apache Arrow) and xorq… | ||
| What We Expect from XAI - A scientist’s experience | Alessandra Costantino | 31m |
Project applies SHAP to remote sensing models for urban resilience, builds maps and a dashboard that surfaces top contributors but requires heavy curation… | ||
| LLM-Based Recommendation Systems | Özge Çinko | 30m |
Shows concrete stacks: LLM embeddings + vector DB (FAISS-style) for offline enrichment, a hybrid pipeline that mixes content and behavioral scores with an… | ||
| The Silent Crash:Why Your RAG Evaluation Metrics Are Lying to You | Hitendri Bomble | 33m |
Dashboards report Cosine Similarity, BLEU, BERTScore and show “90% accurate” while generated answers flip legal roles — nominated beneficiary becomes legal… | ||
| From SQL to Python: Building Data Context for Agents and people | Dmitry Petrov | 25m |
Claims that adding concrete multi-layer context (storage, metadata, dataset, code) moves agents from brittle demos to reliable results, citing 95% accuracy… | ||
| Lightning Talks | 28m | |
Morning session assembles a string of concrete projects and outcomes: Probaviz Streamlit app displays decision boundaries, supports 26 classifiers, parses… | ||
| From Chat-with-PDF to Quiz-Master | Adam Hill | 38m |
Builds a stateful Python app that converts a 130-page PDF into auto-generated multiple-choice and free-text questions using Docling for layout and tables… | ||
| Making tech boring to keep data exciting | Fred O'Loughlin, Kerry Parker, Mark Cottam | 29m |
Climate Policy Radar describes concrete moves: migrating from AWS Step Functions to Prefect and Snowflake, replacing brittle end‑of‑pipeline checks with DBT CI… | ||
| The Clean Energy Graveyard | Damian Bemben | 30m |
The project claims 3,800 cancelled projects in the UK since 2010 and ships an open-source python pipeline & interactive web visualisation that ingests the… | ||
| What Can LLMs Do with Messy Data? | Cedric Clyburn, Andrew Igdal | 30m |
Evaluates what large language models actually do with ResStock-style data using pandas/DuckDB workflows and representative ResStock runs (solar thermal in… | ||
| When Your Dataset Has Blind Spots LLM-Based Data Augmentation | Ophelie Bleu | 25m |
A practical recipe: use LLM synthetic generation to generate 50 reviews for missing classes when there are zero real examples, apply LoRA fine-tuning (example… | ||
| AI-Assisted Creative for Automated Marketing using Python | Matthew Crooks | 30m |
A production pipeline written mostly in Python that automatically renders thousands of platform-specific video and image assets via Adobe After Effects… | ||
| nothing matches | ||