Pablo Castro

Pablo Castro leads the AI Knowledge team in Microsoft’s CoreAI division, building information understanding and retrieval systems for AI applications and agents. Microsoft identifies him as CVP & Distinguished Engineer; his remit includes Foundry IQ, Azure AI Search and Azure Content Understanding. His work concerns the knowledge infrastructure beneath agents: making content usable, retrieving relevant information and preserving access controls.[2]

In his 2026 Foundry IQ explanation, Castro frames production readiness as a combined problem of stability, scale, data access, answer quality, security and ingestion. He describes reusable knowledge bases that multiple agents can access through the Model Context Protocol (MCP), reducing the need to rebuild connectors and retrieval logic for each application. The team’s iterative retrieval engine batches queries, ranks passages semantically and caches tokens on the server to reduce redundant consumption across multi-turn conversations.[2]

Two architectural judgments stand out in Castro’s account: “Security belongs at the data layer, not approximated in application code” and “Ingestion quality sets the ceiling on retrieval quality.” He describes permission synchronization and sensitivity-label controls intended to keep enterprise policies intact as content reaches agents. For ingestion, he explains preview capabilities that preserve document layout and convert diagrams, charts and scanned images into meaningful text, so retrieval can draw on more than raw prose.[2]

At the 2026 AI Engineer World’s Fair, Castro also discussed the interaction between people and agents. In Richard MacManus’s July 1 account for Latent Space, Castro presented Microsoft Foundry as an “AI app and agent factory” and argued that a “learning loop” occurs when people and agents work together.[1]

Castro’s earlier Azure Search work supplies context for this focus on understanding and retrieving information. Before his current AI Knowledge role, he helped shape a search service around customer needs and publicly explained how AI enrichment could make heterogeneous content searchable.[5][3][2]

Castro, Liam Cavanagh and Eugene Shvets undertook a 13-week TechStars experiment to test whether a Microsoft team could operate as a lean startup. A November 2015 Microsoft feature, published roughly two years after their graduation, identifies Castro as partner director of software engineering on Azure Search. He recalled their central lesson: “a startup is a machine that searches for customer value above everything else.”[5]

Customer interviews changed the team’s scope. Its initial plan covered virtually any search-as-a-service scenario, but a generic platform did not address specific customer needs. The team narrowed its focus to catalogs and ecommerce, user-generated content and social networks, and line-of-business applications. “Scarcity forced us to make good choices,” Castro said. He also linked customer understanding to engineering practice: DevOps helped the team understand customer workloads, while continuous delivery put enhancements in customers’ hands every few days.[5]

In his May 7, 2018 Cognitive Search announcement, written on behalf of the Content Search and Intelligence team, Castro explained an ingest–enrich–explore architecture. Ingestion extracted content from sources including PDFs and Office documents, with incremental processing for changes. Composable AI skills added annotations through OCR, entity recognition and image analysis; webhook-based custom skills allowed developers to extend the pipeline. Original content and enrichment annotations then entered a search index for keyword search, structured queries and faceted navigation.[3]