Enterprises adopting AI tooling — whether for developer productivity (vibe coding) or AI-powered services — lack centralized control over LLM usage. Token spending is untracked, prompt injection risks go undetected (a 200 OK no longer means the request was safe), and there is no standard governance layer across teams and applications. Traditional observability is insufficient: security and compliance teams need to know was sent to an LLM, , and whether the response was trustworthy.

Provide a reference architecture on STACKIT (using STACKIT Model Serving and SKE) that deploys a central, open source LLM gateway all developers and applications route their standardized API requests through. The gateway enforces token budgets, logs requests and responses for auditability, and detects prompt injection attempts — giving enterprises the control and observability layer AI workloads require.