--- title: "Lab: talks, workflows and Zeitgeist" description: "Conference talks, workflows from competitions to follow step by step, and Zeitgeist, a magazine run by AI agents." canonical: https://fuerst.ai/en/lab lang: en --- *Lab* # Tried, shown, explained. On stage I show how automation and AI agents work. Here you find my talks, workflows from competitions and Zeitgeist, my experiment with a magazine run by AI agents. ![Florian Fürst speaking at Berlin AI Day 2026 in front of a full room](/img/buehne-berlin-ai-day-wide.webp) *Talks* ## Stages *Coming up* - **11 November 2026, Mannheim:** talk on Zeitgeist at the Claude Code Meetup hosted by Accenture - **25 and 26 November 2026:** mentor at an AI bootcamp for a private equity firm, on behalf of Pantomeno - **Berlin AI Day 2026:** “From Spreadsheet to Solution”, the Goethe-Institut case, from an Excel spreadsheet to a portal - **No-Code Week 2026:** a sovereign RAG agent, entirely without US providers - **Expert video for BIEG Hessen:** “Artificial intelligence: building and using agents”, from simple automation to AI agent (in German). [Video on YouTube](https://www.youtube.com/watch?v=ARMbuNWSziI) - **n8n Meetup Rhein-Main, Wiesbaden:** several times since 2025, on changing topics - **AI Week 2025:** an AI agent in practice ![Florian Fürst in front of a full room at Berlin AI Day 2026](/img/buehne-berlin-ai-day.webp) ![Florian Fürst speaking at No-Code Week 2026](/img/buehne-no-code-week.webp) ![Florian Fürst presenting his school email workflow at the n8n Meetup Rhein-Main in Wiesbaden](/img/buehne-n8n-meetup-schulmails.webp) ![Florian Fürst showing an n8n workflow at AI Week 2025](/img/buehne-ai-week.webp) Looking for a speaker for your event? [Write to me](/en/contact). *From the workshop* ## Workflows to look at. I only show client projects with permission. I built these workflows for talks, competitions and challenges, so I can show them in full. ### Refund Tracker Second place at Flow4Gold at the n8n conference In The Loop 2026: three AI agents follow returns all the way to the refund. They read the drop-off receipt, check delivery with DHL, and look at the bank account and PayPal to see whether the money has arrived. If it has not, they suggest a friendly reminder to the retailer. It only goes out once I approve it. *Capture the receipt* 1. **Receipt uploaded** (Trigger: Google Drive). A photo or PDF of the drop-off receipt lands in a folder. 2. **Read the receipt** (Step: Pixtral, Mistral, Adacor). The image becomes text, and the text becomes carrier, tracking number, recipient and date. *Expect the refund* 3. **Research the refund** (AI agent: Gmail, Bank account, PayPal). Finds the order confirmation and the expected amount. Delivery expected after one week, refund after two. 4. **Add a row** (Step: Data table). One row per return, with amount and deadlines. *Check and act* 5. **DHL agent** (AI agent: DHL API). Checks every day whether the return has been delivered, and records it. 6. **Refund agent** (AI agent: Bank account, PayPal, Gmail). Checks bank account and PayPal. If the money is there, the case is closed. If the shop announces a date, it moves the deadline. 7. **Approval** (Human check: Gmail). The agent proposes a friendly reminder. It only goes out once I approve. 8. **Reminder to the shop** (Step: Web search, Gmail). Looks up the shop’s service address and sends the reminder. Simplified view of the n8n workflow. Refund Tracker, 2nd place Flow4Gold, n8n In The Loop 2026. ### Sovereign RAG agent For my class at No-Code Week Frankfurt 2026: an AI agent answers questions from your own documents. The files live in Nextcloud, the Supabase vector database is self-hosted, both on Hetzner servers. The models come from the IONOS AI Model Hub. Entirely without US providers. ![Florian Fürst giving his class at No-Code Week Frankfurt 2026](/img/buehne-no-code-week.webp) *Take in knowledge* 1. **Schedule** (Trigger: n8n). Starts on its own at regular intervals. 2. **Fetch the files** (Step: Nextcloud, Hetzner). Downloads the PDFs from a Nextcloud folder. 3. **Extract and split the text** (Step: n8n). Takes the text out of the PDF and splits it into sections. 4. **Store in the vector database** (Step: Supabase, IONOS). Each section is stored as a vector, so it can be found by meaning. *Answer questions* 5. **Question in the chat** (Trigger: Chat). Someone asks a question. 6. **AI agent** (AI agent: IONOS, Memory). Understands the question, remembers the conversation and decides when to search. 7. **Find matching sections** (Step: Supabase, IONOS). Finds the sections that match the question. The agent answers on that basis. All in Germany and the EU: Servers at Hetzner, models via the IONOS AI Model Hub, Nextcloud and Supabase self-hosted. No US provider. Simplified view of the n8n workflow. Sovereign RAG agent, No-Code Week Frankfurt 2026. ### School and after-school emails Built for my own family, a winner in n8n's #FrAIday challenge in April 2026 and presented at the n8n Meetup Rhein-Main in May 2026: an AI agent reads the emails from school and after-school care, attachments included, puts dates in the calendar and tasks in the to-do list, and sends me a summary. The prize was an n8n hoodie, and I wore it at the meetup. ![Florian Fürst in the n8n hoodie he won, at the n8n Meetup Rhein-Main](/img/florian-n8n-hoodie.webp) *Incoming* 1. **Email from school or after-school care** (Trigger: Gmail). Certain senders start the workflow. *Attachments* 2. **Save attachments** (Step: Google Drive). Every attachment is stored. 3. **Analyse attachments** (Step: Gemini). Reads the documents, so the agent knows what they say. *AI agent* 4. **AI agent** (AI agent: Claude). Reads the email and attachments, checks the calendar and decides what to add. *Result (depending on the case)* 5. **Events** (Step: Google Calendar). Creates events in the family calendar. 6. **Tasks** (Step: Google Tasks). Creates tasks in the to-do list. 7. **Summary** (Step: Gmail). Sends a short summary by email. Simplified view of the n8n workflow. School and after-school emails, n8n #FrAIday Challenge 2026. ### Support inbox: Inbox Inferno For an n8n community challenge in March 2026: an agent sorts incoming requests by topic. Specialist agents for pricing, security and setup answer only from the documentation provided. Anything they cannot answer with confidence goes to a person. The models run in Adacor's private cloud. *Incoming* 1. **Request arrives** (Trigger: Webhook). An email to support arrives as a record. *Triage* 2. **Triage agent** (AI agent: Qwen, Adacor). Assigns the request to a topic, following the escalation rules. *Specialist agents (depending on the case)* 3. **Pricing** (AI agent: Documentation). Answers only from the pricing and product documents. 4. **Security** (AI agent: Documentation). Answers only with approved responses on security and data protection. 5. **Setup** (AI agent: Documentation). Answers from the product and integration documentation. 6. **Escalation** (Human check: Rules). Whatever cannot be answered with confidence goes to a person, with a polite handover. 7. **Spam and misdirected mail** (Step). Sorted out, no reply. *Reply* 8. **Reply goes back** (Step: Webhook, Gmail). Topic and reply go back, as a record and by email. Simplified view of the n8n workflow. Inbox Inferno, n8n community challenge, March 2026. *Zeitgeist · Experiment* ## A magazine run by AI agents Zeitgeist is my agentic magazine about AI and what it means for people, and an experiment. AI agents monitor sources, research, check every number against the source and write analyses in German and English every day. I set the rules, the checks and the cost limits. - **5 desks** and the n8n Lab - **3 to 6 articles** a day, in German and English - **More than 30 sources** approved, from research to regulators - **$100 a month** cap for the language models ### How it is built *Build* - **Claude Code** develops, tests, deploys. Not part of daily operation. - **Git repository** Workflows as code, prompts and rules, database migrations, theme. *Steering · AI agents* - **Editor-in-chief** daily overview at 7:50 am, review on Mondays, approvals by email. - **Watchdog · error alert** at least 3 articles a day, every error by email. *Operation · n8n on Elestio (Hetzner, Falkenstein) · daily at 4:30 am* - **Sources** more than 30 approved: RSS, lists, Firecrawl. - **Research** Key statements with a verbatim quote. (Sonnet) - **Desk leads** Syntheses first, then 5 desks select. (Sonnet) - **Primary source** finds the original behind retellings. (Sonnet) - **Counterpoint** looks for a sourced opposing view. (Sonnet) - **Newsroom** German and English, with a self-check list. (Opus) - **Standards check** Numbers, evidence, logic, DE = EN. Approves. (Opus) - **Rework** fixes findings, once. (findings, check again) - **Publisher** publishes both languages. - **Title card** - **Timeline** - **Expert request** Their context appears afterwards. - **Operations database (Postgres)** evidence files · tasks and check findings · approvals · own login per agent, only the check approves. Budget lock: at most $100 of model costs a month, every task reserves in advance. *Outside* - **WordPress** zeitgeist.fuerst.ai, German and English, custom theme. - **Anthropic API** Claude Opus 5.5 and Claude Sonnet 5.5. - **Web and mail** Firecrawl, Apify, zeitgeist@fuerst.ai. - **LinkedIn agent** 3 carousels a week. *prepared, waiting for access* - **Requests to sources** open questions by email. *decided, not built* **A human as publisher:** sets the editorial line, sources and budget · decides on approvals and proposals by email · articles go out without prior review, experts add context afterwards Agentic first: agents do the work, rules set the limits, a human is the publisher. Articles are published without a human reviewing them first. Experts put selected articles in context afterwards. Decisions that matter are made by a human. As an experiment, Zeitgeist is not finished: I am testing how far agents can do journalistic work, and I learn from every mistake. *Coming up* - **11 November 2026, Mannheim:** talk on Zeitgeist at the Claude Code Meetup hosted by Accenture [Read Zeitgeist](https://zeitgeist.fuerst.ai/) *First call* ## What would your project be? Tell me in 30 minutes where things get stuck. [Book a first call](/en/contact#termin) ![Florian Fürst listening to a question from the audience at Berlin AI Day 2026](/img/florian-berlin-ai-day-nachdenklich-closing.webp)