← All speakers

Utkarsh Shrivastava

Utkarsh Shrivastava leads engineering for enterprise AI at Netflix, where his organization builds enterprise search and Netflix’s agentic suite. His teams help employees find information across the company, answer questions spanning multiple systems, and carry out work with AI. His current focus is making those capabilities work together as more teams build their own tools: how agents retrieve information, choose between overlapping tools, and decide which actions they can take independently.

On this page

Building shared AI capabilities at Netflix

Utkarsh’s leadership at Netflix has spanned customer-service systems, intelligence engineering for enterprise functions, and data platforms. He led the transformation of customer-service systems through generative AI, led the data warehouse team and expanded its scope to support a data lake, and built a leadership team across locations, including Netflix’s Warsaw engineering hub.

His enterprise-AI work addresses a coordination problem created by faster development. Individual product teams can build their own integrations, but a shared assistant must make useful choices across all of them. Different tools may expose different versions of the same entity or appear capable of answering the same question. Connecting more tools therefore increases both the assistant’s reach and its opportunities to choose incorrectly.

Utkarsh’s work on shared assistants includes several related mechanisms:

  • Evaluating tool selection. His organization uses an internal benchmark drawn from real employee queries to evaluate routing between overlapping tools. This tests whether the assistant selects a suitable tool for the employee’s question.
  • Certifying connectors and assigning ownership. Independently built integrations need to work within a common assistant. Connector certification accompanies decisions about which responsibilities remain with individual teams and which belong to the shared platform. Ownership matters when one team builds on another’s data: the data-owning team may discover that dependency only when an evaluation fails.
  • Classifying agent actions. His work includes distinguishing actions an agent can take independently from those that need additional controls, alongside the separate problem of retrieving information and selecting tools.

For employees in functions such as Finance and Talent, the benefit is being able to connect information in one place that previously required approaching several teams. Utkarsh’s platform work concerns the engineering decisions needed to support that experience as the number of participating teams and tools grows.

Yahoo Mail: experimentation and desktop integration

Before Netflix, Utkarsh led engineering teams across Yahoo Mail’s mobile and desktop products. He helped establish its growth team and led experimentation work.

Yahoo Mail’s Windows 10 application illustrates a concrete engineering tradeoff from that period. The team wrapped its existing web application in a desktop app, preserving its JavaScript implementation and existing features while adding Windows capabilities. Utkarsh explained how a background task called WinJS APIs to update unread-mail counts on Live Tiles and notification badges. Reusing the web application let the team add a visible signal of new mail outside the browser without rebuilding the mail client.

Utkarsh is also one of six named inventors on Yahoo’s onboarding feature cues patent, filed in November 2015 and granted in November 2020. The invention uses user information and cue rules to prioritize and schedule introductions to product features. It addresses both selection and timing: which feature to introduce to a particular user, and when to introduce it. His broader work includes named inventorship on several patents at Yahoo and Netflix.