KI FÜR MÜNCHEN AI adoption. Done right.

AI systems you can prove.

KI für München is not a consultancy that delivers slide decks. Behind the brand is an AI platform engineer with the DNA of a senior test and quality engineer: over 15 years of making stochastic systems testable and provable — applied to AI agents today. What we recommend runs in production with us first.

PROFILE / 01

The person behind it

Stephan Walkowiak

ROLE
AI Platform Engineer & Managing Director
FOUNDATION
Senior test & quality engineer (15+ years)
FOCUS
Self-hosted multi-agent systems in production
LOCATION
Munich

I build self-hosted multi-agent systems that run in production — not in a prototype. My approach: lock down the stochastic part of an LLM and make the workflow around it reproducible, verifiable and compliance-proof.

What sets me apart is more than 15 years of making stochastic systems testable and provable — test automation, the V-model, ISTQB. I apply exactly that craft to AI agents today. In over 3,000 hours with Claude, Codex and Gemini I built my own AI infrastructure self-taught and systematically linked it with my testing and developer know-how — data-sovereign and EU-AI-Act-/GDPR-compliant by design, self-hosted instead of cloud lock-in.

For decades I have worked end-to-end and on my own responsibility: I take full ownership of topics, stay closely embedded in the overall plan and communicate proactively. As an entrepreneur I plan, steer and decide ahead of time — so goals are reached quickly.

This experience, this insight and this knowledge I offer to companies that want to go the same way: analyse processes, identify value, implement, support. All from a single source.

CORE COMPETENCIES / 02

Four layers, one through-line

From agent orchestration to compliance at runtime — every layer rests on the same foundation: building systems you can trust.

K-01

Agentic Systems & LLM Engineering

The stochastic core locked down, the workflow around it provable.

Multi-agent orchestration (LangGraph, LangChain, supervisor pattern)Human-in-the-loopdeterministic hardening of reproducible agentsRAGStructured Output (Pydantic)prompt-injection defenceMCP & API integrationautomated quality evaluation of AI answers

K-02

Local LLM & MLOps

AI on your own hardware — measurable and cost-controlled.

Local inference on AMD ROCm (llama.cpp, GGUF/quantisation)Model selection (Qwen3, DeepSeek-R1, Mistral, BGE-M3)Observability & token/cost tracking (Langfuse, OpenTelemetry)reproducible benchmarks

K-03

Quality & Lifecycle Engineering — the foundation

Over 15 years of making stochastic systems testable and provable.

Test automation (pytest, Robot Framework, Jenkins/CI)Application Lifecycle ManagementV-modelintegration, performance & storage qualificationroot-cause debuggingISTQB Foundation (Test Manager in preparation)

K-04

Governance, Infrastructure & Compliance

Compliance not as an add-on, but as a runtime property.

EU AI ActGDPR / PII redaction as a runtime hookdata sovereignty / self-hostingDocker/Compose, LXDPostgreSQL/pgvector, Neo4j/Graphiti, RedisFastAPIsecrets management (Infisical)TÜV-certified AI Officer & AI Coordinator

SYSTEMS / 03

In production, not in a prototype

Three running systems that show our approach — each with governance gates, test hardening and clear handovers to humans.

S-01

Agentic CI triage & test automation

Large Munich technology company · camera systems

2024 – today
  • Agentic CI triage system as a composable skill chain: red test run → duplicate check → evidence enrichment → classification → fix routing — each step with a clear read/write boundary.
  • Runtime governance hooks: GDPR PII redaction (no real names to the LLM provider), test gate, skill guard. Terminal actions only with human approval.
  • The V-model's requirement-to-test flow fully automated; in production across Jira, ReportPortal, Jenkins and GitLab.

S-02

Self-hosted AI & e-commerce platform

ByteNubes GmbH

since 2020
  • Production stack with 50+ Docker containers / 26 Compose projects live (including our own platform paramaniac.shop) on Saleor & ERPNext.
  • RAG chatbot & voice agent: Claude API, LiveKit, Whisper (STT), Piper (TTS), Chatwoot, WhatsApp.
  • Omnichannel integration (WhatsApp, Instagram, TikTok, Meta) plus Google/Meta Ads APIs; observability with Grafana/Loki.

S-03

Local-inference agent platform

In-house development

2025 – today
  • Hierarchical “agent company” (supervisor → departments → workers, human-in-the-loop) on AMD Ryzen AI Max+ / Radeon (ROCm) — fully local.
  • Knowledge graph & memory: pgvector for fast retrieval, temporal knowledge (“what held when”) via Neo4j/Graphiti.
  • Reference architecture for EU-AI-Act-compliant, cloud-free AI.

CAREER / 04

Certified, not claimed

Whoever sells governance must be able to prove it. Qualifications publicly verifiable — the path there over two decades of integration and test.

since 2021

Managing Director · ByteNubes GmbH

Brand “KI für München” since 2026. Full entrepreneurial responsibility: strategy, clients, delivery.

since 2011

Test & AI engineer (external) · Large Munich technology company · camera systems

Test automation, HW/SW integration and storage qualification; today agentic CI automation.

2008 – 2010

Post-production & system administration

Post-production handling as well as servers and editing systems.

2007 – 2019

Media & camera production · freelance

TV post-production, later camera and media production as a second pillar — replaced by building the e-commerce business from 2020.

2001 – 2007

IT & network engineering · self-employed

Networks, phone systems and security systems — early practice in systems integration and administration.

METHOD / 05

What we recommend, we use ourselves

Our own organisation runs fully AI-driven. Every recommendation we make carries production load with us first — across five layers.

  1. E-01

    Sales & offers

    Enquiries, quotes and client communication run AI-assisted.

  2. E-02

    Development

    Code assistance, review automation and documentation in daily use.

  3. E-03

    Test & quality

    Test-case derivation and automated regression — our own craft.

  4. E-04

    Pipelines & operations

    CI/CD with AI gates, monitoring and automated finding assessment.

  5. E-05

    Administration

    Reporting, document management and planning — AI-led instead of hand-maintained.

That's the difference

Consultants recommend what they have read. Practitioners recommend what runs for them. If a tool does not survive our own evaluation, it never reaches your operation.

PRINCIPLES / 06

How you can measure us

W-01

Data sovereignty

Self-hosted first. Your data, your knowledge, your models stay under your control — cloud only where it is demonstrably the better choice.

W-02

Deterministic, not surprising

We lock down the stochastic part of an LLM and make the workflow around it reproducible: defined gates, clear handovers to humans, every system documented and auditable.

W-03

Service first

We are a small firm with big ambition: the client we have won stays. Because we are reachable, we deliver, and we stick with it.

Away from work: paraglider and ultralight pilot, photography and cameras — the closeness to precise technology stays, even in free time.

CONTACT / 07

Let's get to know each other

A conversation says more than any website. Tell us where your company stands — we'll tell you honestly what we would do.