Rolling out AI guardrails across dev, product and business

Scope

The fourth separate mandate for the same banking client — each engagement won on the results of the previous one: CI/CD platform, AppOps enablement, DORA compliance automation, and now the secure foundation for AI-assisted work in a BaFin-regulated environment.

• Building a central LiteLLM model gateway as the single, governed access point to LLMs — model routing, budgets and auditable usage instead of scattered API keys.
• Developing MCP servers and a company skills library, so agent workflows connect to the bank's real systems and every solved process stays reusable.
• Engineering a secure environment for developers: devcontainer sandboxes with a firewall — controlled network egress and scoped credentials instead of laptop-wide access.
• Engineering guardrails as code: permission allowlists, hooks and approval gates — versioned, reviewed and tested like infrastructure, deterministic instead of trusting the model.
• Rolling out AI-assisted work beyond engineering: developers, product owners and business roles work in one governed setup instead of ungoverned individual tools.

Technologies: Claude Code, LiteLLM, MCP, Devcontainers, GitLab

Key Skills Demonstrated

• AI Governance & Guardrails: Designing deterministic controls for AI-assisted work — policies, approval gates and audit trails that pass an enterprise security review.
• Platform Engineering for AI: Building the gateway, sandboxing and egress layer that turns AI pilots into approved production tooling.
• Compliance Engineering: AI tooling engineered to hold up in a BaFin-regulated environment — continuing the same audit-first approach as the DORA automation mandate.
• Cross-Role Enablement: Bringing engineering, product and business roles onto one governed setup — with adoption measured, not assumed.
• FinOps for AI: Applying Kubernetes cost discipline (kubecost) to tokens — budgets, model routing and cost per team through the gateway.
• Change Management: A versioned skills library that keeps improving the setup after the rollout.