Model Context Protocol Experts in Germany

• Generative AI products for corporate clients, owned from strategy through specification to production.
• Central strategy, local configuration: multi-tenant AI assistant for occupational pension schemes (bAV), delivered as an interactive avatar with text and voice path. Three tenants run on one codebase, each with its own conversation guide, while the knowledge base, guardrails and escalation paths stay central
• Versioned, AI-ready knowledge base composed into a tenant-agnostic voice context and tenant-specific text prompts — the configuration layer that keeps local adaptation from forking the product
• Conversational design: answer limits, scope and off-topic handling, anti-hallucination rules, escalation and lead handover to human advisors
• Five eval suites as a quality gate before any prompt or model change (anti-hallucination, LLM-as-judge failure modes, multi-turn consistency, voice KPIs, action vocabulary with confusion matrix); user test with 10 testers (Hamburg, 07/2026) drove the rework from alpha to beta
• Coordinated external developers, compliance and client stakeholders; GDPR-compliant EU stack, IDD-compliant, EU AI Act classification documented
• Second product line: white-label social media generator for consultancy chilli mind (CH/DE) — one codebase, per-client branding and configuration
• Results: 239+ deployments and a pilot with corporate customers · 108+ deployments for the white-label product · repeatable pattern for multi-tenant AI products in a regulated environment

• Claude Code subagents, MCP, Pydantic V2, pytest, bandit

• Designed and shipped 3 specialized agents that hand work down a line: a research agent writes a cited implementation spec, a coding agent builds the modular code and its tests, a review agent ranks findings by severity and applies the fixes. Each handoff is a structured document, so no stage depends on another agent's context window.

• Connected the research agent to an academic-research MCP server (Semantic Scholar, ArXiv, Hugging Face Hub, citation snowballing) so every reference traces to a tool result rather than the model. Gated commits behind ruff, mypy, pytest and bandit, required human sign-off before installs and commits, and persisted session state on disk so long runs survive a context reset.

• Built a scalable nearshore IT developer hub (Croatia, Czech Republic, Poland) as an independent company through a BOT model (Build – Operate – Transfer)
• Identified, selected, and managed full-service agencies; introduced management and control mechanisms, including KPIs, SLAs, and regular service reviews
• Prepared and reviewed data processing agreements and framework contracts in coordination with Legal & Compliance; integrated regulatory requirements (including KRITIS) into process design
• Advised on cloud vs. on-premise strategies, data storage, and authorization concepts; supported Procurement with tendering and service provider evaluations
• Managed change and process harmonization between internal teams and nearshore partners; reported to executive management, CFO, and CIO

Result: Scalable IT developer hub with an audit-ready governance model, reduced operating costs, and accelerated product development.

• Repaired a broken “ScrumBan” process, then established a pure Kanban system; increased output in the Kanban flow by 22% within four weeks

• Increased team autonomy and decision-making ability by implementing new decision strategies; resulting in up to 25% better outcomes

• Redesigned retrospectives (including one-to-one coaching and workshops), which led to consistent implementation of the resulting action items

• Intensive coaching of Product Owners (POs) to develop and support “Empowered Teams”, alongside leadership development to place agile frameworks and methods in a realistic context (“de-illusioning”)

• Supported change management processes to promote an agile company culture among management and teams, improving internal communication to increase transparency and effectiveness in agile processes

Project: Defect Generation on Test-Bench Images of Metal Surfaces Environment: Automated Visual Inspection (AVI), Metallurgy & Manufacturing

• Objective & Implementation: Designed, architected, and trained Generative Adversarial Networks (Pix2PixHD / SPADE) for image-to-image transformation. Targeted generation of synthetic material defects (e.g., cracks, inclusions, scale) on rough metal surfaces under real test-bench lighting conditions for privacy-compliant and efficient dataset expansion (data augmentation).
• Technical Design: Implemented robust Generative AI and computer vision pipelines in Python and PyTorch. Used semantic segmentation approaches for mask-controlled defect synthesis and subsequent evaluation with EfficientDet object detection models.
• Business Impact: Massive dataset upscaling (10x) without time-consuming and costly physical test-bench runs, while significantly improving the detection performance of automated inspection systems.

Technologies & Skills Used: Python | PyTorch | SPADE | Pix2PixHD | EfficientDet | Machine Learning | Semantic Segmentation | Computer Vision

Stabilization of an Azure/.NET landscape in live operation.

• Architecture, DevOps, and operational readiness; technical decisions under time pressure
• Azure DevOps, monitoring, ETL/ELT, cloud security, FinOps, and data-mesh-related topics

Technologies: Azure DevOps, .NET, CI/CD, monitoring, FinOps

Remote freelance role focused on Agentic AI strategy, secure application patterns, and reusable agentic workflows for a government agency.

• Development and implementation of an open source Agentic AI strategy for a government agency, with a focus on GDPR, security, and self hosted solutions
• Development of reusable agentic workflows and mini applications that enable non technical employees to solve business problems independently
• Implementation of internal business applications with Single Sign On (SSO) and Azure PostgreSQL integration on Hetzner Linux servers
• Implementation of nine mini applications with Single Sign On (SSO) and Azure PostgreSQL integration on Hetzner Linux servers
• Techstack: Python, Nextjs, Typescript, Streamlit, Anthropic SDK (Claude), Azure, Linux Ubuntu, PostgreSQL, MS SQL, Angular, Authentik

• Fully on-premises LLM document-examination platform for a compliance-critical banking domain: agentic LangGraph pipeline with deterministic verification, every AI judgment structured and source-anchored; ~960 automated tests, zero data egress
• GPU throughput engineering (quantized serving, speculative decoding, prefix caching): 9.5x extraction speed-up, 500+ multi-document case files per day on a single A100
• AI-native EDI/EDIFACT integration platform (~116k LOC Java 25 / Spring Boot 4, 1,900+ tests): LLM-drafted partner mappings machine-verified before go-live (DFDL conformance, field-coverage checks, dry runs), ~99.5% byte match on real customer files — replacing weeks of manual mapping per partner

• This project involved extraction of intelligence data to support report writing for a company that provides geopolitical, global, commercial intelligence. The data have been gathered from a number of resources (interview transcripts, online data, internal documents), and then a knowledge base has been build from it. This was the basis of a complex RAG system, that was evaluated against a golden dataset. Agents have been used to find out the contradicting intelligence, the statements supporting each other, and to store back the generated knowledge.

Used: Python, RAG, LangGraph, LangChain, deepeval, MCP

• Recurring problem in my own AI-assisted projects: requirements analysis, use cases, and architecture decisions can be created quickly with AI support, but remain difficult to follow and scattered across Markdown files – knowledge is lost as soon as it is no longer in the context window. arknet turns requirements engineering and architecture knowledge into structured, verifiable data instead of plain text: requirements, use cases, and architecture decisions form a consistently linked knowledge graph, traceable from requirement to architecture decision – queryable by both people and AI agents. Technically based on RDF/OWL and a custom MCP server.

• Result: Working MCP daemon, Docker image published automatically to GHCR, nine hexagonal modules, eleven ADRs (including an Open-Core licensing model). Requirements engineering and Ubiquitous Language hexagons are active. Public as a Community Edition under Apache-2.0 since 07/2026 (github.com/kogn-io/arknet), together with the Claude Code plugin and GHCR image; Open-Core model.

• Label: Java, Maven, RDF, RDF4J, OWL, SPARQL, Model Context Protocol, Spring AI, Docker, GitHub, Git, Claude Code, Obsidian, DDD, Hexagonal Architecture, ArchUnit, JUnit, AssertJ, Interface Development, Software Architecture, Continuous Integration, Knowledge Management

• Ride-sharing and bus platform, live and fully functional. Backend with Spring Boot 4.1 on Java 21, PostgreSQL with Flyway, and Testcontainers integration tests. Hosted in my own AWS account (ECS Fargate, ALB, ECR, IAM Least-Privilege) with CI/CD via GitHub Actions and OIDC federation without static credentials. Development fully AI-supported with Claude Code, including custom skills and project-specific memory. Spring Boot · Java 21 · PostgreSQL · Flyway · Docker · AWS ECS/ALB/ECR · CI/CD · GitHub Actions · Claude Code

Workflow-automation SaaS for operations teams (Berlin, 120 people); full-time freelance engagement reporting to the CEO: an initial 12-month interim mandate, extended twice through the AI build-out; owned product for one squad and coached the other product managers on process.

• Led generative AI (LLM) integration into the core product: from LLM-powered steps to natural-language workflow authoring and step-level automation suggestions, plus AI-managed dynamic workflows, shipped behind eval gates with human-in-the-loop fallbacks: AI-drafted workflows grew to 31% of all new workflows, and median time-to-first-workflow fell from 3 days to 4 hours.
• Packaged the AI capabilities as a usage-based add-on priced on executed automation steps, working with sales and marketing on positioning: ~€800K added ARR in the first year, and adopting accounts churned 1.8 pp less.
• Owned the roadmap end to end: replaced feature-request-driven quarterly planning with an outcome-based rolling roadmap built on quarterly bets and explicit kill criteria, presented monthly to the executive team and quarterly to the board.
• Rebuilt the product-management operating system: weekly customer-discovery cadence incl. workshop facilitation, RFC/decision-doc reviews and a single quarterly metrics narrative; coached four product managers, one promoted to senior during the engagement.
• Closed the engagement as scoped: hired and onboarded the permanent VP Product, handed over the process playbook and roadmap, and exited on schedule in June 2026.

• Team leadership: Prioritization and coordination of four cross-functional teams.
• Platform strategy: Development and implementation of strategies to optimize existing IT platforms.
• Stakeholder management: Active management of expectations and communication with internal and external stakeholders.
• Program and innovation management: Prioritization and coordination of cross-department projects as well as innovation initiatives.
• Product Owner consulting: Advising Product Owners with a focus on product development and continuous product improvement.
• Organizational development: Improving communication and decision-making structures across all organizational levels.
• Change management: Implementing best-practice change management methods to ensure continuous optimization and innovation.
• Quality assurance: Ensuring high quality standards in processes, services, and deliverables.

Seitenkumpel — agents build websites for trade businesses, unattended. Own product, live.

• Agents research public company data and build complete websites from it, with nobody watching
• A validation layer makes sure extraction errors fail loudly instead of passing quietly
• Acquisition runs through a postcard funnel with a screenshot and a QR code, subscription model from 79 euros a month
• Result: several hundred websites built, running unattended
• Honest limit: there are no paying subscriptions yet — the funnel is built, the revenue is not there

Stack: agent workflows built directly without a framework, Claude and OpenAI APIs, TypeScript, Node.js, PostgreSQL, Cloudflare Workers, web scraping, data enrichment