“10 % AI, 100 % software engineering.” As intelligence moves on-device — for privacy, latency, regulation, cost and sovereignty — the cloud’s one-line-API shortcut disappears, and software architecture moves back to center stage.
This session turns the “AI goes local” thesis into concrete architecture. Using a real, fully local reference system — a privacy-preserving agent built from several small, specialized models on a single laptop — Christian reads a local-first Gen AI system as an architecture problem, not a model problem. Along the way: the surprisingly large deterministic core, where a regex, a small classifier or plain SQL beats any model; models as versioned, swappable engineering assets; and the local-first / cloud-optional boundary, where confidence-based routing keeps data on the machine until a human says otherwise.
The throughline: models change, hardware changes — the patterns, and your data, stay.