Quality Engineer · AI Architect · Frankfurt am Main
I build AI automation that survives production.
- Twenty years of quality engineering, three years of AI architecture. The combination is the point.
- Five projects on this page, each with a picture, a result and, where possible, open evidence.
Michael Boiman
I come from quality engineering. That is why I trust no green checkmark.
- I build the systems myself and then run them myself.
- I work with open standards, because anyone can read them.
- Whatever I claim, I back with something you can open.
The AI agent answering on this page runs on a machine in my study.
Working together
What companies bring me in for.
As a project, a review or a workshop. One email is enough to start.
Project · AI automation
An orchestrator that all of my work runs through.
Challenge
Development across many repos, several clients and a dozen tools. Switching context cost me hours every day.
Solution
A central system of addressable agents, a registry across every project and a skill tree for each tool. The agents query each other over the open A2A protocol, and a gateway serves the same agents to MCP clients.
Result
From incident analysis through board upkeep to meeting minutes, everything runs through one entry point. The generic framework behind it is open source under MIT, while client data stays in the private layer.
github.com/bks-lab/open-bridge · MITClone it and look inside
$ gh repo clone bks-lab/open-bridge
$ cd open-bridge && head -1 LICENSE
MIT License
$ ls rules/ | grep -E 'guard|safety'
promote-safety.md
push-guard.md
$ ls agents/ | grep gateway
_gateway
Commands run against the open repo on 21 August 2026.
Project · Voice AI
A voice bot that names real open slots on the phone.
Challenge
A practice loses calls because nobody can pick up during treatment. A language model, however, must never invent an appointment.
Solution
The model runs the conversation, the open slots come live from the booking portal. It pencils in and never books. Consent comes before the first substantive sentence.
Result
The caller hears a short holding line while the answer finishes in the background. The bot says date and time as words, because speech synthesis swallows individual digits.
Voice channel at the clientnot public · The line belongs to the client, the number stays private.
I need an appointment next week
One moment, let me check
On Tuesday the eleventh, at half past ten
Recorded in test operation, reproduced in substance.
Project · Process automation
Read email, sort it, post it into the ERP, with nobody in between.
Challenge
Incoming mail was read by hand, classified and typed into SAP. That cost time every day and produced transcription errors.
Solution
Microsoft Graph fetches the mail, a language model classifies it, Azure Functions process it further and hand the result to SAP over RFC.
Result
Classification runs end to end and the dashboard shows every category over time. When the assignment is uncertain, the mail goes to a human instead of into the wrong case.
Email classification at the clientnot public · The system runs at the client, and it has no public address.
Project · Enterprise integration
Invoices that have to arrive across Europe.
Challenge
Outgoing and incoming invoices had to move in both directions between SAP Business ByDesign and the European Peppol network, in production, and without losing a single document.
Solution
Two routes under the open EN 16931 standard. Inbound over two parallel paths, by webhook and from a mailbox, both ending in the same processing layer.
Result
A forensic check showed that accepted and delivered are two different things. Half the sample had never arrived, and the cause sat in master data rather than in code. The check itself belongs to the client and stays confidential.
EN 16931 · Peppol BIS 3.0Open standards, independently readable
Project · Quality engineering
Release decisions from data instead of gut feeling.
Challenge
Go or no-go decisions cost hours of collecting numbers from the test tools of several teams and several environments.
Solution
One continuous pipeline pulls every source together and shows availability, response times and open incidents side by side on a dashboard.
Result
During a legacy migration the validation ran around the clock in dual operation: old and new route under load at the same time, with every deviation reported by the machine rather than a person.
github.com/bks-lab/experimental-django-sla-dashboardThe dashboard in the picture is public
How I work with AI
The machine works up to the draft. Approving is not its job.
- A human moves the card, and that starts the run.
- Each stage starts fresh, without the context of the previous one.
- The reviewer sees only the artefacts, never the implementer's reasoning.
- Even changes to the system itself pass through the same gate.
To me that is the difference between AI you can deploy and AI you can demo.
rules/learning-autonomy.mdThe rule is public in the repo
Principles
Four rules that have held up for me.
One Bug Policy
After every fix we establish which level of the test pipeline should have caught it. That exact test then gets written.
Logging before Debugging
If an analysis needed the debugger, logging was missing. It ships together with the fix.
Machines do not tire
No human tests a million combinations. A machine that never sleeps does.
Measurable or not at all
Automation without measurement is only faster guessing.
Live on this page
Ask my agent instead of taking my word.
- Probes the deliberately narrow knowledge source.
- Probes whether the agents really reach each other.
- Probes the calendar mirror: times only, no title, no place.
Three questions at its edges: what it knows, who it reaches, what it gives away about appointments.
mboiman.bks-lab.com/.well-known/agent-card.jsonIts agent card, open it directly
not checked yet
- Name
- …
- Protocol
- …
- Address
- …
- Skills
- …
Track record
Where I learned this.
The short version. Every position, certificate and talk is on the CV.
- since 09/2025Senior Quality Engineer · Test Automation and MonitoringTÜV Süd
- since 06/2025LLM Infrastructure Architect & Automation EngineerBKS · AI Research & Development
- 07/2024-01/2025AI-driven Automated QA Environment for Energy InfrastructureAkkuSwap Startup
- 01/2024-04/2025Technical Lead & AI Automation ArchitectBKS
- 04/2023-04/2024AI Engineering Lead & ML Solutions ArchitectBKS on behalf of a client in the sustainability sector
- 08/2021-05/2025Platform Quality Architect (CI/CD, Monitoring, Automation)DVAG
- 01/2017-05/2021DevOps Quality Lead & Dashboard ArchitectDB Vertrieb
- 2000-2005Diploma in Computer Science (UAS)Cologne University of Applied Sciences, Gummersbach
Next step
Michael Boiman
- Write to me about what is coming up. Two sentences on the task are enough.
- Everything on this page can be opened or is marked confidential.
mboiman@gmail.comLinkedInGitHubopen-bridgeBKS-Lab
last built on 22 August 2026Stamped at build time, not typed.
Classic CV · PDF · Deutsch · Privacy · Legal Notice