Zurich Language AI meetup #21

Update: We’re fully booked and running a waitlist.

Get ready for the 21st edition of our community event for language AI and NLP experts and enthusiasts – with a NEW LOCATION at Riffraff Kino/Bar!

🧠 3 straight-to-the-point talks by experienced experts
🥂 Casual networking over drinks and snacks
📍 NEW location: Riffraff Kino/Bar, Neugasse 57, 8005 Zurich
🎟 Capacity: 100 seats

Schedule

6:30 pm – Doors open and registration
7:00 pm – Presentations

🔹 Overcorrection: Why LLMs Can't Just Proofread – and How to Fix It
Heejin Do, Postdoctoral Fellow at ETH AI Center

LLMs often over-edit text, sacrificing original meaning for fluency. This talk introduces PoCO, a two-stage method that harnesses 'intentional overcorrection' to catch more errors while using smaller models to restore accuracy and precision.

🔹 Why Translation AI Fails (Even Though It Works)
Christian Weih-Sum, Managing Director at foryouandyourcustomers

The quality of modern translation AI is no longer the main issue. In practice, deployments fall short not because of models, but because of processes, roles, and decision-making. This impulse explores the non-technical challenges that decide whether translation AI becomes a competitive advantage—or another missed opportunity.

🔹 Privacy vs Utility: The Anonymisation Tradeoff in Language AI Workflows
Michael Anthony Mazourik, Founder and CEO at AI4Privacy

Every AI workflow leaks data it shouldn't. But most anonymisation tools either destroy context or miss entities entirely. To tackle this, AI4privacy built p5y, an open-source privacy framework modelled on i18n, that treats privacy as a language-level translation task. The talk covers reversible masking, synthetic identity g