@pydatalondon
https://london.pydata.org We run a monthly meetup and host the annual PyData London conference. PyData is an educational program of NumFOCUS, helping our community share ideas and learn from each other. https://www.meetup.com/PyData-London-Meetup/
⚡️ Lightning: Mapping the International PyData Community featuring Web Scraping and Data Wrangling - @hevansdev.bsky.social
Hugh loves maps! But Meetup’s map of PyData events is not great.
So he made a better one: circle size for group and green for newly scheduled meet-ups.
⚡️Lightning talk: Agentic, Agentic AI for Personalized Care — Lessons and Challenges from Holisticare.io - Mojtaba Kargar
Cost efficiency levers with Hyperpod. And we even get nice Grafana stats!
A blueprint of questions to ask before you provision your servers.
Considerations accessing GPUs and run distributed training. AWS has a lot of options! The right choice depends on whether you’re experimenting, scheduling a fine tune or running a full multi-week training job.
Our next talk is From Zero to HyperPod: Cutting Infrastructure Complexity for Distributed Model Training on AWS - Anton Nazaruk
We’ve just seen a 2 minute demo of calculations that would have taken hours before!
And there is useful explainability: here’s a plot of the first DMD mode overlaid on a map.
It also uses dynamic mode decomposition for SVD for spatial dimensionality, similar to PCA. But noise can be a big problem—though luckily we can fix with a post processing step.
SVD-ROM uses Dask extensively. Dask has two algorithms for SVD: TSQR (tall-and-skinny) exact an randomised approximate. The randomised results are almost as accurate and MUCH faster.
SVD-ROM implements Reduced Order Modeling (ROM) algorithms based on the Singular Value Decomposition (SVD)
It’s really useful for high dimensional dynamical systems like weather or climate data.
Our first talk tonight is SVD-ROM: Reduced Order Modeling of huge arrays using the Singular Value Decomposition - David Salvador-Jasin
Excited to be back for another PyData London meetup! 🥳
Slides for Understanding Embedding Space talk
docs.google.com/presentation...
Slides and the GitHub project with all the code that was demonstrated
docs.google.com/presentation...
github.com/pilipolio/ch...
Lighting talk: Local MCP servers for your Local coding Agents (Cursor, Claude, Codex etc) - Prashant Tiwari
MCP systems are distributed and can get out of sync if multiple agents are working at once. Using a shared MCP server simplifies management and helps manage shared resources.
vectors made modern LLMs possible but there are some pitfalls. Here’s what could be done.
Lighting talk: Understanding Embedding Space: Surprises, Pitfalls, and Intuition - Alper Nebi Kanlı
But this matters beyond jokes. Are we missing valuable answers because the models are being too safe?
Lightning talk: Amusingly Abliterated LLMs - Ian Ozsvald
Even if you ask it to be rude, LLMs will tell derivative dad jokes. Can we make an LLM tell good jokes if we remove the guardrails?
It’s definitely rude, but also more original.
Project Architecture and example output snapshot showing the results of querying the symbolic chess endpoint and LLM reasoning.
He uses an agentic flow combining LLMs with symbolic chess engines to generate grounded chess commentary.
Our first talk is Productionising research papers with Python and serverless infrastructure - Chess commentary generation use case @guillaumealla.in
Excited to kick off our 102nd PyData meetup! Jiarui is compering us tonight and as ever we’re hosted by the kind folks at Man AHL.
1. And finally we have a lightning talk on USearch: The Engine Behind Most New Search & RAG Pipelines - Ash Vardanian
USearch might be the most widely used search engine you’ve never heard of, up there with FAISS and Lucene. It even supports semantic search and geospatial indexing!
Our first lightning talk is Mise and easy: managing environments the fun way - Connor Adams. Environment management is very messy across languages. Mise can make your life much easier!
Find pytest-fixture-fixture on GitHub and install via pip or uv
github.com/fferegrino/p...
It works with CV, JSON, JSONL and YAML,
has custom deserialisation, many (many!) out of the box fixtures, and even parameterisation.
Pytest fixtures great but are slightly strange things. They feel a bit too much like magic sometimes.
@feregri.no solution is pytest-fixture-fixture (really!)
Fixtures, for fixtures. Handy fixtures to access your test fixtures from your pytest tests.
Data-driven testing is an approach where test logic stays the same, but the input data and expected outputs come from external sources.
Here’s a Pytest example.
Next tonight we’re hearing about No More Boilerplate: Data-Driven Testing Made Easy from @feregri.no