Bruno Zamberlan

As AI agents began executing processes across different operations and channels, the need emerged to decouple agent logic from client- and channel-specific infrastructure. I designed and built a generic, channel-agnostic platform for executing and orchestrating AI agents, separating ingress, invocation adaptation, cognitive logic, and action delivery. The architecture emerged from real production needs and is now used across multiple operations, allowing the same engineering foundation to be reused in different business contexts.

• > Channel-agnostic architecture enabling the same agent logic to operate across different interfaces and entry points
• > Separation of ingress, invocation adaptation, cognitive logic, and action dispatch
• > Agent and workflow orchestration with explicit state management
• > Integration with conversational channels and HTTP APIs without coupling them to the agent core
• > Event Ledger for execution and platform event traceability
• > Human-in-the-loop support for processes requiring human intervention or validation
• > Observability and monitoring mechanisms for production operation and troubleshooting
• > Architecture reused across different operations and business contexts

A healthcare operation relied on disconnected systems for conversations, scheduling, and internal management, creating information gaps and operational bottlenecks. The solution evolved from data integration and operational visibility to decision support and, later, workflow execution through automation, AI, and conversational systems.

• > ~6K conversations processed per month
• > 12 physicians supported by the operation
• > ~2× estimated operational capacity per team member, from ~5 to ~10 physicians per operations professional
• > Integration of operational and conversational data into daily and weekly intelligence workflows
• > Operational intelligence covering performance, conversion, loss reasons, and next actions
• > Internal agents transforming conversational information into structured tasks
• > Conversational systems integrated with confirmation, pre- and post-service, scheduling, and rescheduling workflows

A real estate auction lead generation application relied on a low-code architecture that did not scale effectively for the volume and complexity of the required data processing. I built and deployed a Python backend to handle this pipeline, integrating auction platforms through web scraping, document processing with OCR, information extraction, and enrichment through external services. The solution was decoupled from the frontend, containerized with Docker, deployed to VPS infrastructure, and load tested to validate its behavior under increased processing volume.

• > Python backend built to move data-intensive processing out of the low-code layer and enable the solution to scale
• > End-to-end automated pipeline from property discovery to structured lead generation
• > Integration with auction platforms through web scraping
• > Unstructured document processing using OCR and information extraction
• > Data enrichment through external service integrations
• > Decoupled architecture between the Python backend and low-code frontend through APIs
• > Docker containerization and deployment to VPS infrastructure
• > Load testing to validate capacity and system behavior under increased processing volume