Portfolio · AI projects
Not experiments. Put into production.
Reference projects with examples, measured outcomes and documentation. Each with its starting point, the decision and the result, the way I would write it in a status report.
Automation cockpit · in production since 11/2025
End-to-end AI process chain
Starting pointRecurring research and document work, done by hand every time and slightly differently every time.
SolutionA chain of n8n, language models, REST APIs and webhooks on self-hosted Docker infrastructure: data intake from 6 interfaces, pre-filtering, AI scoring, research and personalised documents.
- around 80 records per run, approx. 40 % qualified automatically
- a complete document package in about 3 minutes
- runs every 20 minutes, unattended
Quality assurance
Measure, don't trust
Starting pointThe language model rated the quality of its own output at 92 % across the board.
DecisionMeasured instead of trusted: actual requirement coverage ranged from 50 to 100 %. That became a hard limit in code with targeted automatic correction.
- measurable rules enforced in code, not requested from the model
- a block on invented content; anything unproven is flagged as a gap
- automated review workflows check every result before release
- more than 140 automated tests before every release
Vendor and technology decision
Model choice based on data
Starting pointAn analysis of around 1,000 runs showed that one provider rejected every request because of a plan limit, while the fallback quietly absorbed it.
DecisionRoot cause proven via the rate-limit headers, then the model re-chosen per task and tested on 27 real cases against the previous setup.
- free EU model for volume, a stronger model only where quality drops
- criteria: quality, data protection with training opt-out, cost
- automatic failover on outages stays in place
Capstone project of the training
AI in the PMO
Starting pointRisk registers, multi-project overviews and status reports take PMO time that is missing for decisions.
SolutionGDPR-compliant AI agents for risk management, multi-project management and status reporting.
- overall workflow of six sub-workflows, documented and signed off
- handover artefacts: short docs, runbook, checklists, UAT
- human-in-the-loop: approvals stay with people, GDPR from day one
Live on this site
This site and its assistant
SolutionNode back end with a strict content security policy, rate limiting and a server-side token. The assistant runs through a secured n8n workflow; no key ever reaches the browser.
- labelled as AI under Art. 50 of the EU AI Act
- no third-party sources, no tracking, self-hosted fonts
Operations, not a demo
Own platform
SolutionCoolify, n8n, Forgejo, NocoDB and a self-hosted search instance behind Traefik, all on my own hardware.
- versioned, documented deployments
- certificates, backups and health checks
- monitoring with push alerts
- duplicate detection and failover paths