Projects
15 hands-on projects – from running n8n systems with local language models to process analysis, data governance and EU compliance. Anonymised case studies at the level of real company scenarios in the German and EU mid-market.
Answering bookings, menu and allergy questions automatically – local and GDPR-compliant
Industry: restaurant, Berlin
n8nMulti-agentLocal LLMsVoiceGDPR
- Task
- Staff spend hours a day on bookings, menu and allergen questions and confirmations. Allergy information is health data – a special category of personal data under Art. 9(1) GDPR.
- Solution
- 7 n8n workflows: text and voice dialogue, booking rules in code, reminders 2 h before, nightly quality evaluation, monitoring. A custom router fine-tuned locally (Qwen3-4B + LoRA), agents on a local model, personal data masked before every model call, unclear cases handed to a human.
- Result
- Router accuracy 86.8 % (46 of 53 dialogues, measured)
- Routing in 0.25 s instead of 3.89 s (measured)
- Running cost approx. €550/year (model estimate)
- Local is approx. €50/month more expensive than cloud – deliberately, so special-category data under Art. 9 GDPR stays within own infrastructure.
Architecture at a glance
- Guest · messenger
- Notice Art. 13 GDPR · Art. 50 AI Act
- Masking of personal data
- Router · Qwen3-4B + LoRA
- Agents: booking · menu · concierge
- Validator · 1 retry
- Human if unclear
Answering customer questions by voice message – fully local
Industry: retail company
n8nVoice AIwhisper.cppRAGLocal
- Task
- Many repetitive questions about stock and delivery. Customers should ask by voice message, get a spoken answer based on real inventory data and be able to order.
- Solution
- Speech-to-text (whisper.cpp), language model (Gemma), vector search (Qdrant, bge-m3) and text-to-speech (Piper) run locally; n8n orchestrates.
- Prices and stock come from code, not from the model.
- AI disclosure under Art. 50 AI Act, personal-data filter before speech output, watchdog every 10 min.
- Result
- 8-step guided ordering.
- Voice and personal data never leave own infrastructure; no transfer to AI providers in third countries (Art. 44 et seq. GDPR).
- Five code-based quality checks.
Architecture at a glance
- Customer · voice message
- Speech-to-text whisper.cpp (local)
- Guardrails
- Classifier · Gemma (structured)
- Stock: Postgres · FAQ: Qdrant
- Answer from facts (code)
- Personal-data filter
- Text-to-speech Piper (local)
- Staff if unclear
Pre-sorting 400 patient requests a day – routine automated, risk to humans
Industry: clinic network, Berlin
n8nMulti-agentLLM-as-a-judgeHuman-in-the-loopPseudonymisation
- Task
- 9 staff handle 400 requests a day (appointments, prices, lab-result status, complaints). Routine cases should be answered automatically, risky ones handed over to people, and every answer checked independently.
- Solution
- Pseudonymisation (name, phone, e-mail, insurance number, date of birth) before every model call.
- Router and 4 specialist agents; 6 clearly defined hand-over conditions for staff (confirm / edit / replace).
- Every answer checked by a model from another vendor.
- Medical result contents – health data under Art. 9(1) GDPR – never reach the cloud model.
- Result
- Test run with 16 requests (measured): 16 of 16 routed correctly
- 0 missed escalations
- 0 data-protection violations
- 69 % answered without staff (target: 45 %)
Architecture at a glance
- Request · form
- Pseudonymisation
- Check: data removed?
- Router
- Agents: appointment · prices · result status · complaint
- 6 hand-over conditions
- Staff: confirm · edit · replace
- Check by a second model
RAG agent with a knowledge base
Industry: IT service provider
RAGn8nHuman-in-the-loopVector database
- Task
- Make company knowledge usable through an AI agent – on a current, clean knowledge base.
- Solution
- Two connected n8n workflows: a RAG agent with human-in-the-loop approval, plus automatic knowledge base updates with deduplication.
- Result
- An agent system with three outcomes – answer, approval by the legal department, escalation; the lawyers’ answers automatically extend the knowledge base. All 6 test cases passed in live operation.
Architecture at a glance
- Question · messenger
- Search in the handbook (Qdrant)
- Agent with structured output
- A: answer with source
- B: approval by legal
- C: escalation
- Lawyer
- Knowledge base update
Structured AI output system (n8n)
Core technical project
n8nStructured outputSelf-correcting parserError handling
- Task
- Make AI output reliable and machine-readable – even when the model gets it wrong.
- Solution
- An n8n workflow with a self-correcting parser, structured guided reasoning, map-reduce for long texts and global error handling (error trigger).
- Result
- One form for screenshots, PDFs and audio; bug reports with severity, summaries even of long documents (map-reduce). 4 tests passed in live operation; unknown formats do not cause a crash.
Architecture at a glance
- Upload · form
- Switch by file type
- Screenshot · PDF · audio
- Structured output with self-correction
- Map-reduce for long texts
- Sheet · error handling
Finding the right candidates with one chat question – instead of reading CVs one by one
Industry: HR department
n8nRAGPineconeStructured outputAI Act
- Task
- Recruiters read every CV separately and can hardly find suitable profiles in the archive again.
- Solution
- New CVs are captured automatically: 8 key facts go into a sheet and a search database.
- After that, one chat question is enough – e.g. “Who has SAP experience and speaks German?”.
- Deliberately with test data only: AI in candidate selection is high-risk under Annex III AI Act.
- Result
- The archive becomes a searchable talent pool. Instead of reading stacks, the team reviews a shortlist – the decision stays human.
Architecture at a glance
- CV · form
- Text extraction
- Agent: 8 key facts
- Sheet
- Vector database (Pinecone)
- Chat question
- Search by meaning
- Recruiter decides
Mailbox on autopilot: sorting e-mails, checking appointments, preparing replies
Industry: service company with e-mail appointment booking
n8nGmailGoogle CalendarLLM classification
- Task
- A company receives many e-mails every day – enquiries, appointment requests, invitations – and reads, sorts and answers them by hand.
- Solution
- n8n workflow with 54 nodes.
- AI classification and reply draft in the language of the e-mail.
- Calendar check with a concrete alternative slot.
- Attachments filed by client, logging.
- Notifications only during working hours, protection against reply loops.
- Result
- Prototype with a video demo. Routine mail is pre-sorted and prepared – a person reviews and sends.
Architecture at a glance
- New e-mail
- AI classification
- Calendar: slot free?
- Reply draft in the thread
- Attachments by client · log
- Person reviews and sends
Capturing ID data without retyping – with staff approval
Industry: real-estate agency
n8nVision LLMHuman-in-the-loop
- Task
- Customer data from an ID scan should reach the sheet without retyping, finally confirmed by a person.
- Solution
- A vision model reads 8 fields with a confidence score.
- Unreadable fields stay empty – no guessing.
- Staff receive the scan and a pre-filled form.
- Result
- Prototype. No more retyping, no invented values – a person confirms.
Architecture at a glance
- Scan · messenger
- Vision model: 8 fields with confidence
- Draft in the sheet
- Staff review
- Confirmed data
FinOps model and AI roadmap for a clinic network
Industry: clinic network, Berlin
TCOROIScenariosRoadmapRisk register
- Task
- The board has to decide which AI initiatives to fund, what they cost, when they pay back and what happens in a bad case.
- Solution
- 18-sheet Excel model: token costs with separate input/output prices, TCO incl. hidden costs, 3 scenarios, sensitivity, risk register, 12-month roadmap; board pitch deck. Health data (Art. 9 GDPR) processed locally only – although cloud would be cheaper; the extra cost is shown openly.
- Result
- Model estimate, base scenario: investment €129,315, effect €522,460/year, 3-year ROI 229 %, payback in month 13; the conservative scenario (−14 %) is shown openly. Initiative 1 was later built as the “patient service” project.
Architecture at a glance
- Current cost of routine work
- Token costs: input and output separate
- TCO incl. hidden costs
- Effect with conversion factors 0.25–0.60
- 3 scenarios
- Risk register
- 12-month roadmap with gates
EU AI Act compliance package
Industry: staffing agency
EU AI ActRisk classificationDPIAVendor due diligence
- Task
- Classify and safeguard an AI system used for scoring job applicants.
- Solution
- Risk classification under the EU AI Act (high-risk per Annex III), a full three-part DPIA and vendor due diligence for the AI SaaS components.
- Result
- A compliance package with a reasoned approval decision – the core principle being that risk follows the intended use, not the name of the feature.
Architecture at a glance
- Intended use
- Risk class under the AI Act
- Three-part DPIA
- AI vendor review
- Measures with traffic lights
- Reasoned approval decision
AI-optimised process map
Industry: dairy / food manufacturer
Process analysisAS-IS → TO-BERoot-cause analysisAI touchpoints
- Task
- Analyse an inefficient business process and redesign it – using AI where it genuinely adds value.
- Solution
- Full process analysis (AS-IS), root-cause analysis (“5 Whys”) and design of the target process (TO-BE) with AI touchpoints marked deliberately.
- Result
- AI-optimised TO-BE process map including documentation, presentation and a technical talk.
Architecture at a glance
- Current process (AS-IS)
- Root-cause analysis “5 Whys”
- AI touchpoints
- Target process (TO-BE)
- KPIs
Data governance and data quality framework
Industry: dairy / food manufacturer
Data governanceRACIData contractData quality
- Task
- Establish clear ownership and measurable data quality as the foundation for AI projects.
- Solution
- A framework with 12 components: ownership map, RACI matrix, data contract, data lineage and quality dimensions with concrete metrics.
- Result
- A comprehensive framework document (multi-part Excel model) with presentation and role model.
Architecture at a glance
- Ownership map
- RACI matrix
- Data contract
- Data lineage
- Quality dimensions with metrics
AI portfolio canvas – prioritising AI initiatives
Industry: cloud compliance provider
AI strategyPrioritisationROI assessmentExecutive summary
- Task
- Pick the right ideas out of many possible AI initiatives – in a way management can follow.
- Solution
- A portfolio canvas with a weighted prioritisation formula (value / effort / risk) applied to every initiative.
- Result
- A prioritised AI portfolio plus an executive summary in management format (half a page).
Architecture at a glance
- Collect AI ideas
- Value · effort · risk
- Weighted prioritisation
- Portfolio
- Executive summary for management
AI marketing assistants with brand and legal review
Training case based on the Kärcher brand (no commission)
Prompt engineeringGeminiMulti-roleUWG
- Task
- B2C and B2B copy must keep the brand voice; unsupported advertising claims are a legal risk in Germany (UWG).
- Solution
- 3 assistants: creative assistant; a 10-role orchestrator (router → specialist → brand guardian → legal review → compliance); support assistant with an FAQ base.
- Human approval before the final version.
- Result
- 7 prompt versions up to delivery. Own review afterwards: tone checks do not replace fact checks – later projects therefore add a separate source check.
Architecture at a glance
- Briefing
- Router
- Specialist role
- Brand guardian
- Legal review · compliance
- Human approval
Consulting method E0–E8 for AI automation
Own method derived from 15 projects
Process modelGatesGDPRAI Act4 languages
- Task
- Turn individual projects into a repeatable approach for any industry.
- Solution
- 9 phases with approval gates, 7 principles, 12 client documents, adaptations for 6 sectors (SMEs, consulting, public sector, education, healthcare, private); offline generator in 4 languages – client data never leaves the browser.
- Result
- The method used to prepare a pilot project (dental practice).
Architecture at a glance
- 9 phases E0–E8
- Approval gate after each phase
- 7 principles
- 12 client documents
- Generator in 4 languages (offline)