Sudhir Tonse

Sudhir Tonse built and led DoorDash’s Data and AI/ML Platform organization, which started its shared machine learning platform with one job: predicting when a delivery would arrive. That first workload gave his team a way to test the whole path from training a model to serving a prediction in production. His engineering career spans Netflix’s move to the cloud, Uber’s marketplace data systems, and the work of finding failures in AI agents once people start using them.

Keeping Netflix services connected

Sudhir worked at Netscape, Sun Microsystems, and AOL before helping Netflix move from DVDs to streaming and from its own data centers to the cloud. He was responsible for the Netflix Cloud Platform as the company moved to independently deployable services. Those services still had to reach one another when individual instances slowed down or disappeared. Sudhir and Allen Wang jointly announced Ribbon, an open source client library Netflix built to distribute requests across instances of a service. Sudhir also co-invented a method for introducing delays into a distributed system so engineers could see how the rest of an application behaved when one part slowed down.

Building around the first delivery estimate

At Uber, Sudhir managed Real-time Data Intelligence, the engineering organization responsible for the marketplace data pipeline, the systems that stored and processed its events, and marketplace forecasting. He joined DoorDash in 2019 to build its data platform organization. DoorDash teams already had uses for machine learning and needed shared infrastructure to train, deploy, and operate their models.

Sudhir’s team chose delivery ETA prediction as its first workload. The team responsible for arrival estimates already wanted platform support and could help test the workflow end to end. The platform team began with a shared prediction service, so each model didn’t need a service of its own. Search and recommendations then put different demands on the design: customers can browse many restaurants without placing an order, so those features called for far more predictions than delivery estimates did. The architecture evolved as more teams put it to work, and the organization Sudhir built and led grew to more than 100 people. In his account of the project, he recommends choosing a first customer whose problem exercises the whole system, then letting later workloads show where the design needs to change.

Finding failures that evaluations miss

After DoorDash, Sudhir became vice president of engineering at Galileo, which built tools for evaluating generative AI and monitoring it in production. Cisco completed its acquisition of Galileo in May 2026, bringing that technology into Splunk’s observability portfolio. For AI Engineer Code 2026, Sudhir was listed as Director of Engineering at Splunk, leading Core Platform & Research for AI Observability and Resilience.

The description of Sudhir and Khasimul Asif Mohammad’s joint session on production agent failures starts with a practical problem: evaluations can pass while users encounter failures that only become visible as a pattern across sessions. Running a reasoning model over every session is expensive, and many alerts can point to the same underlying problem. The two speakers’ session description lays out the choices this creates: what to screen cheaply, what warrants closer inspection, and how to group related alerts so teams know what to fix first. It also asks what should earn an agent-proposed fix a merge. A regression test for the newly discovered failure gives the team a way to check the repair and catch the same problem if it returns.