Clara

The problem

Enterprise procurement data and actions live across disconnected systems — Procol, Ariba, Coupa, SAP — each with its own event model and integration contract. Building point-to-point integrations for every new client was slow, brittle, and high-maintenance.

The approach

• Introduced a mediator pattern to decouple producers from platform-specific adapters.
• Used MCP (Model Context Protocol) as a standard integration surface so tools and platforms can plug in consistently.
• Built a dedicated AI orchestration service to decide what to publish, where, and when.
• Added a Long-Term Memory Store for episodic context across sessions so Clara remembers what happened last time.

Architecture

• Producer emits domain events into a central mediator.
• Mediator routes events to platform adapters (Procol, Ariba, Coupa, SAP) and to the AI service.
• AI service reads from and writes episodic memory to the Long-Term Memory Store.
• MCP layer exposes tools and capabilities to the agent without leaking each platform's internals.

Impact

What I learned

• A mediator earns its complexity when the adapter count is >2 and contracts keep changing.
• Episodic memory changes user expectations — users start treating the agent like a colleague, not a search box.

Tech