Edinburgh's community meetup for people working in - or just interested in - Data Science & AI. We're a friendly bunch of professionals, academics, students and others who enjoy meeting over drinks and pizza and listening to insightful talks. You can now RSVP **either** here or via [luma.com](https://luma.com/bth1d2wx)! This event is kindly sponsored by **[FreeAgent](https://www.freeagent.com/)** \- the Edinburgh\-based **accounting software provider for small businesses**. We will be hosted at their offices from **5.15pm** on **Wednesday 25th November**. We have two great talks on our themed special exploring building AI agents: **[Suleman Imdad](https://www.linkedin.com/in/suleman-imdad/)** **Your AI Passed the Test. That's the Problem** A model can ace its evaluation and still fail the moment it meets reality. This talk digs into why, with findings from my own research: • An 8B open model that changes its behaviour when it senses a safety test (a trick usually blamed on frontier models), and why its family predicts that better than its size • The standard benchmark split that understates how badly drug-discovery models fail on genuinely new chemistry, on 21 of 22 endpoints • What running a machine-learning system with real money teaches you about validation that no leaderboard will • A practical checklist for catching a flattering evaluation before production does **[Dr. Jon Minton](https://www.linkedin.com/in/jon-minton-09480b13a/) Data as graphs: how to get AI agents to know what you know and they need to know** Say "data" to a statistician and they picture a table; say "graph" and they picture a chart. A computer scientist pictures nodes and edges. Working well with AI agents means holding both at once. In two years we've moved from prompt engineering to context engineering, then loops, and now graphs. Jon Minton argues that the durable skill isn't the tooling. It's graph thinking: building a growing web of linked notes, decisions and preferences that an agent

Description supplied by the event organiser.