Munich Datageeks Oktober Edition

As always, there will be plenty of time for networking, discussions, and snacks & drinks courtesy of our host.

📍 Event Details

📅 Date: Tuesday, October 27th, 2026
📍 Venue: Smart Village
📫 Address: Ganghoferstraße 66b, 80339 München

🎤 Munich in Motion: Four Data Agents Planning for a Growing City by Juergen Albertsen, Exasol
by Jürgen Albertsen, Exasol

As Munich grows, decisions about public transport, cycling, roadworks and infrastructure are becoming increasingly connected. A construction project can affect transport routes, cycling corridors and entire neighbourhoods, while population growth and changing mobility patterns create new planning challenges.
In this talk, we will build a practical agentic data intelligence prototype using publicly available Munich datasets. We will bring public transport schedules, bicycle counts, weather information, roadworks, population data and city geography together in Exasol Personal.

Four specialized data agents will explore the data from different perspectives. A mobility operations agent will identify potential disruption zones. An urban planning agent will compare population density with transport and cycling infrastructure. A sustainability agent will investigate how weather and seasonality affect cycling demand. Finally, a city leadership agent will combine these findings to highlight areas that may require further attention.

Each agent works with the same governed data foundation, translates natural-language questions into inspectable SQL and presents its findings through a dedicated Dash Server view. The session will also show how ambiguous questions, data limitations and access controls need to be handled when moving from a simple chatbot to a trustworthy data agent.

Rather than presenting a fictional real-time control centre, this talk provides a reproducible example of how agentic data intelligence can support mobility planning and evidence-based decision-making. Attendees will see how data can be transformed into a shared analytical model that they can explore and extend themselves.

About Jürgen Albertsen:
Jürgen is a Product Manager at Exasol responsible for their operational platform and SaaS offering. He has 30 years of experience in various technology and product leadership positions, focusing heavily on platform and cloud technologies.

🎤 Stop Reranking Your Way Out of a Bad Representation
by Edward C. Zimmermann

The standard hybrid retrieval recipe is BM25 plus embeddings, fused with RRF, then a reranker like ColBERT to clean up the mess. Each stage exists to repair damage done by the one before it. That’s affordable in a datacenter. On a laptop, it isn’t.

BM25 is a bag-of-words heuristic: a normalisation patch on term weighting that grew out of the constraints of computers decades ago. It became the mantra because it was convenient, not because it models how documents work. Single-vector embeddings make the opposite mistake: they keep a blurry gist and lose the exact terms. Neither sees the integer columns, dates, and typed fields that real documents carry. And chunking finishes the job by cutting passages that nest, annotations that overlap, quotations that cross editorial boundaries, and claims linked to evidence elsewhere.

We think much of the reranking bill is the cost of compensating for flattened representations. Fix the representation, and you stop paying to repair it.

LLMs are remarkably good at language: they recognise similarity, infer unstated relationships, and synthesise across passages. They shouldn’t be spending that talent reconstructing structure that retrieval threw away. CoreQuarry indexes lexical, semantic, and typed-object constraints in a shared textual coordinate space, drawing on HyTime and Gonnet’s PAT arrays, so every constraint is evaluated against the same source regions.

It’s a graph, but not a conventional one. In a knowledge graph, edges are asserted: someone extracts them, stores them, and hopes they’re complete. It is a textual graph. Here, the relations (before, after, within, contains, overlaps, intersects) fall out of a shared textual space. Two annotations don’t just have a link between them; they occupy, share, or cross the same words. The edges are implicit in the geometry of the text, not bolted onto it.

Vector similarity asks, “what text is semantically related?” Structural search also asks, “where does this occur, what contains it, what overlaps it, what is it connected to, and what belongs with it?”

CoreQuarry provides a queryable structural representation that keeps the spatial character of text. Your LLM stops reasoning over scraps and starts working on selected regions of a structured textual space.

About Edward C. Zimmermann:
Ed is a transdisciplinary technologist, systems architect, and entrepreneur whose career spans the evolution of modern networking and information retrieval. His experience reaches back to the early days of the ARPAnet—when he found himself as a teenager with a shared office in UCLA’s Boelter Hall—through the early 1990s, when he founded one of the first commercial ISPs in Germany and became an early contributor to the nascent World Wide Web.

For the past twenty-plus years, Ed’s research and development has focused on the intersection of text retrieval, data mining, NLP, computer vision, and machine learning. He has led publicly funded R&D projects in partnership with German, EU, and UN organizations, collaborated extensively with national scientific research institutes, and contributed to a wide array of foundational open-source software packages.

Currently, his work centers on developing CoreQuarry, an "off-grid," high-performance hybrid knowledge engine built to ensure data sovereignty. This research into localized hybrid retrieval and hardware-aligned optimization is proudly supported and funded by the NLnet Foundation / NGI-Zero Commo

📋 Agenda

19:00 - Doors open & networking
19:10 - Welcome & introductions
19:15 - Talk #1: Munich in Motion: Four Data Agents Planning for a Growing City by Juergen Albertsen, Exasol
20:00 - Pizza & networking
20:30 - Talk #2: Stop Reranking Your Way Out of a Bad Representation by Edward C. Zimmerman, CoreQuarry
21:15 - Open networking