Skip to content
Henry Rincón

Case studyData Strategy Lead

A data platform for AI agents — Vesting (2023–2025)

The data platform of an AI-agent startup, designed from scratch: data for the agents to operate, and data about the agents so they could be trusted.

  • Microsoft Fabric
  • Data governance
  • Real-time monitoring

By the numbers

  • 12

    clients integrated into the platform

  • 27

    agents built with the core process

  • 23

    agents monitored at once, in real time

  • 11

    stages of the replicable core process

Context

Vesting, an automation-agents startup, was growing without data infrastructure: its AI-agent analytics had nowhere to live and every client data integration was handcrafted.

The challenge

Build from scratch a data ecosystem able to support AI-agent analytics and monitoring in production — with governance from day one and without slowing the startup down.

How I did it

  1. The architecture on Microsoft Fabric

    I designed the ecosystem on Microsoft Fabric: big data, data warehouse and distributed processing over a lakehouse on OneLake, with pipelines that received the n8n flows orchestrating the agents. The key decision was normalizing events at ingestion: sessions, requests, responses, times, states and token cost, the same for all 27 agents.

  2. What is common and what is configurable

    A new agent or client could not require a new table for what already had a name on the platform, and it could bring its own attributes. The common part was the condition for comparing; the configurable part, the condition for not slowing the startup down.

  3. Governance by design

    I defined the governance to integrate client and system data with an event contract validated at ingestion. The 12 clients were isolated in separate workspaces: the access boundary matched the client boundary and did not depend on a well-written filter.

  4. Monitoring agents in production

    I built real-time monitoring in Power BI, with the session as the unit of analysis and four dimensions: operational, usage, economic and functional. The functional one was the hardest: an agent can respond without resolving anything.

  5. The eleven-stage core process

    I structured, documented and validated the core process used to build the 27 agents: from the use case and the level of autonomy to the definition of done, adoption and observable operation. The asset was not one agent, but the ability to build the next one without me.

Impact

  • The data ecosystem became operational from scratch on Microsoft Fabric: 12 clients integrated, 120 tables, 20 GB and 1,000 events a day.
  • 27 agents in the inventory and up to 23 monitored at once, in real time.
  • The replicable core process, in 11 stages, standardized how agents are designed and shipped.
  • The exact bridge between analytics engineering (DP-600) and AI engineering.
What I take with me

The greater a system's capacity to act, the greater the traceability of its data, its behavior and its limits must be.