Every case study is labelled: Live means in real use, Demo means not in real use (a prototype or a study).
hotel-nike-haus-wessels.de
Input
Bot on my agent platform (Python)
Card in Telegram
Signed /api/v1
Payload and cockpit
Operations
Channels
FastAPI and Pipecat
Agent as configuration
Hybrid RAG
Actions and log
Own MCP server
Quality
seniorenbetreuung-lala.de
Visitor on seniorenbetreuung-lala.de
Vercel rewrite /api/chat/*
Web chat API (FastAPI)
Prompt assembly
Language model
Output
infinitylibre.com
Guests in five languages
Server Components
Payload CMS 3 in the same app
Neon Postgres
Hooks and Server Actions
Vercel Cron, daily
Capture
Local storage
AIClient
Next.js API on Vercel
Structured output
App views
Trigger
Headless agent
Own MCP server
Deterministic checks
Human approval
Execution and log
Sources
Ingestion
Supabase Postgres
hybrid_search_chunks
Context builder
Route handler /api/chat
UI and logging
Smaller pieces of work without a case study of their own.
You play a faulty AI that makes reality glitch. Vanilla JavaScript without a build step, GSAP for cascades and effects, a CRT look in CSS, procedural sound via WebAudio, save state in localStorage. The daily anomaly is derived deterministically from the date.
A local multi-agent system, deliberately designed on paper first: 19 architecture decisions as ADRs, C4 diagrams on three levels, an FMEA with failure modes per component, an event bus design with Redis Streams and a Postgres outbox, and a permission model based on capability tokens. Product definition, architecture and repo structure are done; the runtime code is not built yet.