PyData London reposted this
Meetup #76: PyData London 105th meetup Tuesday, 3rd March 2026 (#proofofnetwork) Yusuf Ganiyu from AstraZeneca explored how data engineering is moving from AI as a simple tool to AI as a true teammate. Instead of the usual chatbot flow: ask a question, get a suggestion, copy and paste the code to run it, agentic workflows start with a goal, reason through the problem, take action, review the results, and keep iterating until it’s fixed. The agent consists of a reasoning engine to think, a planner to split tasks up, tools to take action like running queries or checking logs, and memory to keep track of context. These agents can monitor data quality more intelligently, retry minor failures automatically, suggest fixes for bigger issues, and escalate serious problems with full diagnostics attached. Yusuf even showed how an agent can debug pipelines, detect schema changes, and open pull requests, but with guardrails in place. That said, AI still can’t truly understand business context, company politics, or complex trade-offs, which is why maturity matters. So start small, give read-only access to agents first, and remember that your human judgement is still your biggest advantage. Jethro Reeve showcase his brilliant personal project called “Explore the Kingdom” (https://lnkd.in/e-xr9_dU), built using Cursor, which lets you explore the UK’s social and political landscape in a way that feels almost interactive and alive. You can compare house prices, rent, and energy costs across all 650 constituencies to see where you might fit in, dig into the 2024 election results and historical swings, and visualise income, deprivation, and education data. Behind the scenes, he pulled together 15 government data sources and mapped everything to constituency level using clever aggregation methods, fallback strategies, and clear provenance tracking to keep the data trustworthy. He also created synthetic profiles and fast heat maps to make the experience smooth, plus a nowcasting pipeline that pulls from RSS feeds and research to simulate how events might shift trends. Built with a modern web stack and deployed online, the whole system shows how today’s AI coding tools can act almost like a data engineer; browsing data, querying databases, and following defined workflows without opening a single spreadsheet. Latest AI models have crossed an important threshold in capability, but agentic tools are still rough around the edges, so you need strong guardrails and a healthy dose of caution. Stelios Christodoulou showed how to use GitHub Codespaces as a cloud-based development environment to perform data analysis without setting anything up on your own machine. By adding simple configuration files to your project you can create a consistent, ready-to-go workspace for everyone on the team. Comparing with other tools like Google Colab and Kaggle, there is a free option for individuals (up to 60 hours/mo) before moving to straightforward pay-as-you-go pricing.
