One Registry to Rule them All - Sonny Merla, Mauro Luchetti, & Mattia Redaelli, Quantyca
Summary
Sonny Merla (Global Data Science and AI Manager, Amplifon), Mauro Luchetti (AI Center of Excellence Manager, Quantyca), and Mattia Redaelli (AI Engineer, Quantyca) present how Amplifon — the world's largest hearing care company operating across 26 countries with 20,000 employees and 10,000 stores — built an enterprise-grade registry system for MCP servers and A2A agents to bring order to a fragmented, multi-team AI development environment.
The talk opens by framing the core problem: dozens of teams across three continents were independently wiring up AI agent connections, reinventing security models, and spinning up their own infrastructure — producing chaos at scale. In January 2025, Amplifon launched the Amplify program, a global cross-functional initiative structured around three pillars: governance (regulatory alignment, centralized guidelines), platform (certified infrastructure and deployment processes), and factory (reusable, scalable solution development). The program is managed through a control tower (strategic decision-making on security, legal, and use-case prioritization) and a country-level committee (tactical rollout and value delivery).
The technical architecture built by Quantyca consists of four interconnected layers. First, an AI gateway provides unified model access via a single endpoint, enforces authentication through Entra ID integration, manages per-use-case budgets (configurable weekly/monthly), and centralizes auditing and monitoring of all LLM requests and responses across the enterprise. Second, a private MCP registry extends the open-source community MCP registry specification with enterprise metadata: server ownership, deployment environment (dev/test/prod), authentication mechanism, cost attribution linked to the AI gateway budget system, and use-case linkage for impact analysis and auditability. Internal custom servers and pre-approved curated public servers are both catalogued here. Third, an A2A registry is built on the agent card standard, cataloguing all deployed agents with their identity, endpoint, capabilities, supported modalities, and authentication requirements. When an agent is deployed via CI/CD, it automatically publishes its agent card to the registry, making agent development self-documenting and enabling runtime discovery. Fourth, a use-case registry ties everything together — mapping agents and MCP tools to specific business use cases, capturing which AI models are used (supporting maintenance readiness when model lifecycles end), and rendering full object lineage graphs showing connections between use cases, agents, tools, and models.
The platform demo shows the registry UI with a catalog spanning use cases, MCP servers, and A2A agents. A sample use case ("ticket optimization with AI") displays its status, version, linked assets (agents and MCP servers), AI models in use, and lifecycle history. An MCP inspector lets developers connect to any registered server and verify its exposed tools. A2A agent card compatibility checking is similarly available. Widget-based form builders allow developers to generate server.json or agent card definitions without starting from raw JSON.
To standardize development, Quantyca built two GitHub template repositories — one for MCP servers and one for A2A agents. Both blueprints include Docker files, a package manager, FastAPI-based server scaffolding, built-in authentication, cost tracking, and integration with Langfuse for observability, tracing, and evaluation. The A2A blueprint is framework-agnostic (compatible with LangChain, Agno, or any other framework), relying on port/interface contracts rather than enforcing a specific library. CI/CD pipelines triggered on branch tagging automatically publish Docker images to the artifact registry and push metadata (agent card or server.json) to the central registry backend, closing the loop on discoverability.
The talk concludes by summarizing achieved outcomes: a centralized catalog enabling governance visibility across all MCP and A2A assets, full traceability of use cases through their agents, tools, and models, production-ready developer blueprints that eliminate boilerplate security and infrastructure concerns, and automated CI/CD pipelines that keep the registry synchronized with deployed artifacts. The platform is near production deployment at Amplifon, with capabilities still being expanded.
Key Takeaways
- •Build a private MCP registry as an enterprise extension of the open-source community spec, enriching each server with ownership, environment, authentication model, cost attribution, and use-case linkage to enable impact analysis and auditability.
- •Implement an A2A registry based on the agent card standard and wire it to CI/CD so that every deployed agent auto-publishes its card — making agent development self-documenting without manual catalog maintenance.
- •Create a use-case registry that links agents, MCP tools, and AI models to business use cases, then render object lineage graphs so teams can instantly see what is affected when a model is deprecated or a component has an outage.
- •Deploy an AI gateway as the single entry point for all LLM calls to centralize authentication (Entra ID), per-use-case budget enforcement, and audit logging of all requests and responses.
- •Use GitHub template repositories (blueprints) for both MCP servers and A2A agents that pre-wire Docker, FastAPI, authentication, cost tracking, and Langfuse observability — so developers can focus entirely on business logic.
- •Make the A2A agent blueprint framework-agnostic by defining port/interface contracts rather than requiring LangChain, Agno, or any specific framework, preserving team autonomy while enforcing interoperability.
- •Integrate registry metadata publishing directly into CI/CD tagging workflows so that the registry stays synchronized with production deployments automatically, not through manual updates.
- •Include an MCP inspector and A2A agent card compatibility checker in the registry UI so developers can verify tool exposure and protocol compliance without leaving the platform.
- •Treat AI governance as a three-pillar program (governance, platform, factory) with a central control tower for strategy and a country-level committee for tactical rollout to balance centralized standards with decentralized execution.
Transcript
What happens when you have dozen of teams across three continents all building AI agents, each one wiring up their own connections, reinventing their own security model, deploying their own infrastructures? You get chaos. Hi, I'm Sonny Merla, Global Data Science and AI Manager at Amplifon, and I'm here today with Mauro Luchetti, AI Center of Excellence Manager, and Mattia Redaelli, AI Engineer at Quantica, the team that design and build the technical solution that we are about to describe. Today...
Related Sessions
