PyData London Meetup at Birkbeck

Venue: Room MAL B33, Birkbeck, University of London, Malet St, London WC1E 7HX Please note: 1. 🚨🚨🚨 This event is in a new venue in Birkbeck University. 🚨🚨🚨 2.🚨🚨🚨 A valid photo ID is required by building security. 🚨🚨🚨 3. This event follows the [NumFOCUS Code of Conduct](https://numfocus.org/code-of-conduct). Please familiarise yourself with it before attending.

If your RSVP status says "You're going" you will be able to get in. No need to show your RSVP confirmation when signing in. If you can no longer make it, please unRSVP as soon as possible.

Code of Conduct: This event follows the NumFOCUS Code of Conduct. Please get in touch with the organisers with any questions or concerns.

As always, there will be free food and drinks, generously provided by our host, Man Group.

Main Talks

1. [Sultan Al Awar](https://www.linkedin.com/in/sultan-al-awar/) and [Ksenia Shishkanova](https://www.linkedin.com/in/ksenia-shishkanova-5b35b5119/)- From PoC to Production: Building Scalable Agentic Applications As more companies explore the potential of Gen AI, many hit the same challenge: turning a cool demo into something that actually works in production and brings real value. In production, teams quickly run into issues like unreliable tool execution, missing evaluation baselines, prompt regressions, unclear governance, poor observability, weak versioning practices, and vendor/model lock-in. This talk focuses on how to move beyond prototyping and build agentic systems that are measurable, reliable, and safe to deploy. We’ll walk through the evolution of agent design, starting with prompting and RAG, then progressing to tool-using agents and multi-agent orchestration, and highlight what breaks at each stage. You’ll learn practical AgentOps patterns including how to version agents as AI assets, evaluate performance using LLM-as-a-Judge, build golden datasets with human feedback, and implement end-to-end observability with tracing and online/offline monitoring using MLflow. The session will conclude with a production checklist and a reference architecture for deploying agents with model-agnostic routing, guardrails, payload logging, and rate limiting (with examples from real-world platforms such as Databricks Unity AI Gateway). This talk is aimed at data scientists, data engineers, and AI practitioners building LLM applications, and requires only basic familiarity with Gen AI concepts like LLMs and prompting. 2. [Kavit Tolia](https://www.linkedin.com/in/kavittolia/) -- Listening to black holes with missing data Future space-based detectors will record gravitational waves from merging black holes. But turning a noisy, gappy signal into a statement about the black holes that made it needs Bayesian inference, and real data breaks the usual likelihood-based methods. This talk is a practical tour of simulation-based inference: instead of writing down a likelihood, you train a neural network on simulated signal-and-parameter pairs to learn the posterior directly. Using black-hole "chirp mass" inference as a worked example, I'll compare neural posterior estimation against classic MCMC and pull out three lessons that travel well beyond astrophysics: why your input representation matters as much as your model, why missing data hurts based on where it falls rather than how much is gone, and why a posterior sitting right on the true value can still be wrong. Built entirely in Python (PyCBC, sbi, emcee). No physics background needed.

Lightning Talks

1. [Richard Hickling](https://www.linkedin.com/in/rthickling/)- Making Python Faster

--------------------------------

Logistics

• Doors open at 6.30 pm (get there early as you'll need to sign in with building security). • Talks start at 7:00 pm, with drinks afterwards from 9:00 pm at a nearby pub (TBC).

We have reduced capacity for this event, but there will be plenty of people to discuss data science questions with. Please unRSVP in good time if you realise you can't make it. We're limited by…