CAIML #44

CAIML #44 happened on September 8, 2026, at TH Köln and was sponsored by KölnBusiness.

Agenda

18h30 Open Doors

19h00 Welcome & Intro

Community poll: The most telling number is the smallest one: only 10% picked "not applicable". In a room of around 80 people, Talk-to-Your-Data has stopped being a demo topic — almost everyone is building it. Where it hurts is another matter. Text-to-SQL accuracy and tool integration together draw just 28%, while semantics and data foundations account for 62%. Getting a model to write a query is no longer the hard part; telling it what a metric means is — which is exactly where Tim's talk picked up.

19h15 Tim Voßmerbäumer (Data & AI Consultant at Scalefree): Talk to Your Data: A Semantic-Layer Architecture for Reliable Analytics Agents

An LLM can produce valid SQL over a warehouse schema and still answer the wrong question, because the schema does not carry the organizational meaning of a metric. This talk presents an agentic system over Google Ads, Google Analytics 4, and Google Search Console data, modeled in Data Vault 2 on BigQuery, in which a governed semantic layer serves as the agent’s knowledge interface. The agent answers along two paths: questions that map to a predefined metric go through the dbt Semantic Layer, using definitions that already exist rather than ones the model writes. Questions no predefined metric covers take an exploratory path, where the agent writes its own SQL under a rule-based guard. The agent may also decline when the data cannot support an answer. Without that option it does not fail visibly, but produces SQL that passes every structural check, fabricates the rows, and reports them faithfully.

19h50 Ivan Herreros (Senior AI Consultant | Machine Learning Engineer at inovex): From ChatGPT to Agentic AI: The Common Thread of AI Development

How is Agentic AI actually progressing, how will it keep reshaping day-to-day work, and is any of it really impossible to predict? To answer that, we trace the evolution of LLM-based applications: from natural-language dialogue to the delegation of digital workflows to agents that act autonomously over ever-longer time horizons. The common thread: viewed with some distance, most of the steps (from RAG through MCP to Claude Code and Agent Harnesses) were not just predictable but predicted. Drawing on almost three years of building conversational and agentic AI platforms, we show which patterns recur and what actually sits behind the buzzwords. Whoever understands this underlying logic can situate new technological developments proactively, instead of reacting to every new hype cycle.

20h20 Networking with food and drinks provided by KölnBusiness

Join the discussion on CAIML #44 here.