Iris

Iris is a personal knowledge platform that connects world news, emerging technology and practical learning. It exists because staying informed and actually building something are usually two separate activities, and I wanted one path: understand what is happening, work out why it matters, then learn how to apply it.

• My Briefing. Topic-based news with a focus on Germany, Europe, world affairs and Guatemala.
• Tech Observatory. Developments that matter to AI engineers, data scientists and technology consultants: models and developer tools, but also hardware, robotics and the environmental cost of computing.
• Builder's Lab. Learning paths with documentation, courses, exercises and compared resources, so curiosity ends in a skill rather than a bookmark.

• An interactive interface with light and dark themes and topic navigation.
• Beginner AI engineering learning paths, and visual explanations of agent workflows.
• Source-linked news editions, collected through a separate scheduled workflow.

• The news editions are not yet connected to the interface. That is the next step.
• The agent workflows currently run as educational simulations, not as a production backend.
• The planned backend will coordinate research, analysis, verification and editorial work, and make the process behind each briefing visible rather than hiding it behind a finished article.
• Public deployment is still pending.

Berlin Family Scout

What do we do with a toddler today? I kept answering that question badly at nine on a Saturday morning, with a phone in one hand and a child in the other. So I designed the app I actually wanted: one button that takes where we are, the time, the weather and which children are with us, and returns one recommendation and two backups, read out loud, with somewhere to eat afterwards.

• Places before events. Playgrounds, libraries, farms and indoor play areas are there every day. Events are the fragile, expensive data. Starting with places means the app is useful in week one and cheap to keep alive.
• Open data instead of scraping. OpenStreetMap, the Berlin open data portal, kulturdaten.berlin, the German Weather Service through Bright Sky, and BVG journey times. The commercial listing sites are used to discover venues, never copied, and always linked back to.
• Provenance on every fact. Changing table, stroller access, how much of the place is covered when it rains: each value records where it came from, when it was last checked, and unknown stays unknown. The model is never allowed to fill a gap with a plausible guess.
• Scoring in code, language in the model. Distances, opening hours, age bands and the weather rules are deterministic and testable. The AI does the messy reading of venue pages and writes the one sentence that explains the pick.
• An evaluation set from day one. Thirty places I know by heart, hand checked, run automatically on every prompt change so a change that makes the answers worse is caught before it ships.
• Built to be asked, not opened. Voice in and out, a Friday evening plan for the weekend, and an MCP server so I can ask it through Claude on my phone instead of opening an app at all.