Skip to content

Innovation lab

MVPs, AI prototypes, agentic solutions and digital products. This is where I explore ideas before turning them into structured initiatives.

Reference architecture: enterprise agentic AI

An agent's credibility doesn't come from the model but from its design: what it can query, what it can do, who approves it and how it is evaluated.

  1. Layer 1Interaction
    • User or process
    • Channel (chat, API, workflow)
  2. Layer 2Orchestration
    • Step planning
    • Context management
    • Task memory
  3. Layer 3Knowledge and tools
    • RAG over authorized sources
    • Internal APIs
    • Data queries
  4. Layer 4Control
    • Per-tool permissions
    • Human approval
    • Action limits
  5. Layer 5Operations
    • Traces and audit
    • Quality evaluation
    • Observability and cost
The control layer separates what the agent may propose from what it may execute. Every action is logged, and sensitive actions require human approval.

Projects in the lab

Have an initiative that needs to move from idea to a working solution?

Tell me about the context and the goal. We'll work out together what is worth building, why, and how to start.

Let's talk