Enterprise AI Platform & Agent Infrastructure for FUNKE Mediengruppe

Working closely with Chief AI Officer Dr. Paul Elvers and Executive Board Member Stephan Thurm, I co-architected and built the enterprise AI and agent infrastructure for FUNKE Mediengruppe, one of Germany's largest media organizations. Operating across editorial, business, and operational divisions, the platform serves 5,000+ active users with a 92% monthly retention rate, generating an estimated €2M+ in annual productivity value (featured in an official Google Cloud Customer Success Story).

To turn scattered AI tools into a governed, company-wide capability, we built the platform around an architectural paradigm called "One platform, three doors":

1. The Workbench (Open WebUI): An accessible interface where newsrooms, business teams, HR, and legal can switch between models and create tailored internal assistants.
2. The Gateway (FUNKE AI API via LiteLLM): A central routing and policy layer that "right-sizes" spend—directing high-volume queries to fast Gemini Flash models and complex tasks to frontier reasoning models—processing over 40 billion tokens with granular budget and SSO controls.
3. The Agent Layer (Google ADK & Gemini Engine): A managed, serverless runtime powering autonomous agents, shared memory, and sandboxed execution environments.

My work focused on hands-on development across the agent framework, tools, sandboxes, and enterprise rollouts:

• Multi-Agent Orchestration & Core Agents: Engineered production agent templates (chat & batch) with automated CI/CD and Terraform IaC modules. Built the multi-agent Business Development Agent Team (supervisor agent, specialized subagents, shared state synchronization, and custom Open WebUI rendering functions), alongside production agents like the Synthetic / Deepfake Image Detection Agent (Cloud Run GPU inference with computer vision benchmarks, Sightengine, and Copyleaks integrations) and the Deep Research Agent (a 5-stage pipeline cutting hours of editorial research to under 10 minutes).
• Model Context Protocol (MCP) Ecosystem: Designed, deployed, and secured a registry of 20+ enterprise MCP servers packaged in Artifact Registry. Connectors include internal documentation (MPP360), Germany's national news agency (dpa), regional editorial databots, statistical data (Destatis), Periscope/Poliscope, and Slack integration via FastMCP, with full OAuth 2.1 / OIDC token delegation and secret management.
• Secure Execution Sandboxes & Isolation: Architected and built an MCP server abstraction layer on top of Google Cloud's sandbox solution (Agent Engine runtime). Implemented user management with a custom OAuth delegation flow so individual employees could provision and interact with secure, isolated sandboxes with shell access and computer use capabilities restricted exclusively to them.
• Model Governance & Intelligent Right-Sizing: Configured centralized LiteLLM gateway governance routing over 40 billion tokens across Gemini Flash, Claude Opus 4.7/5.5, and Sonnet 4.6. Implemented granular SSO user budgets, token rate limiting, and automated credential rotation.
• Enterprise Client Rollouts & Enablement: Led the corporate rollout of modern AI desktop and CLI clients across business units and executive leadership—integrating Claude Desktop, Claude Code, OpenCode, and OpenWork into existing IT systems. Tailored client-specific configurations, permissions, and security guardrails, and built automated build-time bi-lingual documentation (DE/EN) for the internal FUNKE AI Developer Portal.

← Back to Projects