Evaluating some ideas

I came into this course with some potential project ideas to choose from.

Greg’s checklist sounds simple: find a real problem, make sure you have data, give the AI a concrete job, be honest about your blockers. Applying that to your own ideas is harder than it looks.

Here are the ones I worked through — all flying around somewhere between health and mobility:

Each one taught me something. About where public data actually exists in Germany. About what a clean agentic loop could look like. About the difference between a product with soul and a product with a real AI job.

What looked like hesitation was actually learning to think in AI product terms.

But in the end, I landed on Policy Assistant — an AI agent that helps product teams assess the legal risk of their AI features and design appropriate guardrails. It came from a real personal frustration. The EU AI Act is public, dense, and complex — perfect RAG material. The agentic loop is clean. The user is my own professional community.

Not the flashiest idea. But I learn two things at the same time: creating an AI product and navigating the legal challenges of developing one.