Belgium NLP Meetup

Thirty meetups! It feels surreal to think back to the very first Belgium NLP Meetup in 2016, when our opening talk posed the question: is there a market for NLP? Safe to say, almost ten years later we have our answer.

Our thirtieth event will take place on Thursday June 4th in the offices of Ravical in Ghent, an AI company that discovers revenue opportunities in documents and helps full-service firms act on them.

As usual, doors open at 7pm for pizza, talks are planned between 7.30pm and 9pm, and after that there's more time for networking and drinks.

The first two talks of the evening are already confirmed. The final presentation will be announced in the coming weeks. Stay tuned!

The Context Window Is Not Enough: Why Real Work is Hard for AI Agents
Kasper Rutten (Archer and Summit)
Language models often begin from a blank slate, while humans rarely do. Every professional task carries a hidden bundle of context: prior conversations, organizational history, domain conventions, emotional stakes, implicit constraints, and memories of what has and has not worked before. This talk explores that contextuality of work through the lens of software engineering in a complex environment, alternating between lived practice and what neuroscience tells us about memory, attention, and recall. From there, we map the same example onto modern AI systems, distinguishing between pre-training, post-training, and in-context learning, and showing where today’s techniques for context management help or fall short. The final part reflects on what this means for the future of AI agents. Rather than assuming instant, economy-wide automation, the talk argues for a more nuanced timeline: one where progress depends less on raw intelligence alone, and more on how well systems can acquire, preserve, and act on context.

Hermit: an experiment in autonomous applications
Ben Schrauwen
Ben sees a new category of software emerging, autonomous applications, and built a small open-source project, Hermit, to explore that idea. Most AI products add an assistant to a static app: fixed workflows, fixed UI, fixed logic. An autonomous application is different. The agent is not a feature inside the product. It is the operator of the product. It owns the workflow, shapes the data model, generates the interface, writes code and skills, and evolves the system over time. The agent is the primary interface: you work through it, but you can still inspect the underlying state directly, while it generates the views needed to make that state legible.

Safe natural-language-to-code execution to level up your AI product
Nico Gelders (Ravical)
Generative AI tools struggle when users ask data-heavy questions, and function calling alone doesn't offer a reliable solution. Nico will show how compiling natural language into sandboxed TypeScript instead (using typed tool schemas as grammar) enables parallel API chaining and produces high-quality, reproducible artefacts. Expect practical lessons from running NL-to-code pipelines in production, including where this pattern works and where it breaks down.