We are running the KI Sommer Camp again — this year as a focused two-day intensive at PAZ Central, Niederdorfstrasse 77 in Zürich, on the 2nd and 3rd of July 2026. In person, small group, real briefs on the table. After the three-day 2025 edition, we cut the format down on purpose: two days is enough to build one working system, and a working system is the only thing that survives the trip back to your own desk on Monday.

Here is the honest version of what most AEC AI talk leaves out. The interesting part is not the model. It is the dependency graph around it — where your project data lives, what cleans it, which tool calls the model, where the output lands, and which step quietly breaks when one input changes. The camp is built to make that graph visible and then make it run.

What the two days cover

We work from the same backbone our 2025 cohort used (documented in our KI Sommer Camp 2025 writeup): GenAI and LLM fundamentals, a little Python, and Grasshopper for the geometry side. The applied blocks are where it earns its keep — using AI to clean and visualise messy project data, running facade and layout variations, and standing up your own small AI assistant instead of pasting prompts into a browser all day. As in 2025, you work on a live competition brief under our eye, not a toy dataset.

For Archicad users we go further than chat. We show where automation actually belongs — GDL, the Python/JSON API, and the PAZ Archicad Add-Ons (PAZ-BOX, and PAZGPT for steering Archicad by AI chat and voice) — so the AI layer plugs into the model you already maintain rather than becoming a fourth tool you have to babysit. The point is integration, not accumulation.

Atelier: The single most useful exercise of the two days is unglamorous — draw the real dependency graph of one current project. Not the org chart, not the BIM execution plan: the actual flow from brief to data to model to drawing. Most desks find a third single point of failure they did not know they had. That finding is the whole point.

Hack: This Hack teaches you to turn a messy competition brief into structured data with one model call. The medium is runnable Python; the domain is AI / ML. Drop your brief text into brief and ask the model for clean JSON you can pipe straight into a schedule or a Grasshopper input.

import anthropic
client = anthropic.Anthropic()
msg = client.messages.create(
    model="claude-opus-4-8", max_tokens=400,
    messages=[{"role": "user",
        "content": "Return only JSON: every room and its area (m2) from:\n" + brief}])
print(msg.content[0].text)

Three lines of real work, the rest is wiring. Once that JSON exists, the model stops being a novelty and becomes a node in your pipeline.

←TODAY: In 2026, the architect who wins time is not the one with the most AI tabs open, but the one whose project data flows cleanly into one assistant they actually trust. →3012: The Zurich-3012 practice runs on systems its makers can still draw from memory. Fulcrum: A two-day camp matters only because the thing you build in it is small enough to understand and sturdy enough to keep.

Who it is for: architects, BIM and project coordinators, and engineers who are tired of AI demos and want one repeatable workflow. Beginners are welcome — bring a laptop or use one of ours. Full price, schedule and registration live on the 2026 camp page; seats are limited because the small-group, strong-feedback format is the product, exactly as we framed it in our open call for workshop leaders.

If you only do one thing this week before deciding: draw that dependency graph. Then come build it with us.