LLM Practitioners Meetup Karlsruhe #4

Hey there,

it's time for the next LLM Practitioners Meetup!

It will take place on Tuesday, May 12th, 2026, at 6:30 PM, in the SteamWork Coworking Space (Roonstraße 23a, 76137 Karlsruhe). We'll provide food and drinks.

Our schedule (please read the talk descriptions at the bottom):

6:30 pm: Get together
7:00 pm: Session #1: Local LLMs vs. Cloud Computing Fraud: A Year of Lessons and Model Progress (Talk by Jonas Scholz)
7:30 pm: Pizza break
8:00 pm: Session #2: Beyond APIs: LLM-Driven Web Automation for Real-World Processes (Talk by Aaron Röhl)
8:30 pm: Discussions, pizza and drinks (until approx. 10 pm)

Looking forward to seeing you all there!

--------

Talk descriptions:
Talk #1: Local LLMs vs. Cloud Computing Fraud: A Year of Lessons and Model Progress (by Jonas Scholz)
Last year I gave a talk at this meetup about fighting cloud computing fraud with local LLMs. The system worked, sort of. We had a tiny Mistral model, a fragile prompt chain, and a lot of duct tape. It was good enough to ship, not good enough to sleep at night. A year later, almost everything in that talk is wrong!

This talk is the rewrite: what I got wrong the first time, what changed in 2026, and what the system looks like now that small local models are actually good at this.
Plus a live demo: a real phishing sample from production, classified faster and cheaper than the frontier models.

Bonus: also a great talk to attend if you want to learn how to circumvent our detection ;)

Talk #2: Beyond APIs: LLM-Driven Web Automation for Real-World Processes (by Aaron Röhl)
Services without APIs have traditionally been a dead end for scalable integration. Browser automation filled the gap but introduced high maintenance overhead. LLM powered browser agents redefine this approach by interpreting websites, analyzing UI structure, inferring intent, and executing workflows across forms and web applications.
The shift is profound: from fragile Playwright scripts to self-adjusting browser automation.
In a live demo with browser-use, we showcase autonomous browser automation and highlight both its strengths and its current limitations.