How Ankur Gupta is making AI model deployment more reliable

For a machine learning engineer, deploying a model should be a straightforward step toward putting it to work. Behind that request, however, sits a chain of infrastructure systems, each capable of delaying or stopping the rollout. At LinkedIn, Ankur Gupta made those connections the focus of a reliability transformation that changed how teams diagnosed failures, assigned responsibility, and delivered permanent fixes. 

Gupta, a Senior Staff Technical Program Manager specializing in AI infrastructure and distributed systems, was brought in to reset a deployment reliability program that had struggled to achieve sustained improvement. Multiple teams had spent more than a year investigating failures and reviewing progress. Yet engineers deploying models still faced long waits and repeated retries. According to Gupta, the missing piece was a dependable way to connect what customers experienced with the systems and teams responsible for resolving it. 

Gupta presented these themes at the AI Infra Summit in Santa Clara on September 16, 2026, in a Compute Track session titled “What It Takes to Run AI Infrastructure at Billion-User Scale.” The session examined the architectural decisions, GPU efficiency strategies, and coordination practices involved in running AI across shared infrastructure. Its focus on the connections between systems and teams reflects the same challenge at the heart of his deployment reliability work: Making the entire path to production dependable.

The platform at the center of the effort is a control plane: Software that coordinates the steps required to put an AI model into production. It validates a deployment, tests it on a small scale, checks its behavior, and manages a gradual rollout. Completing that sequence depends on systems responsible for computing capacity, networking, traffic routing, and other services. A component could appear healthy in isolation while the overall deployment remained stuck. 

Gupta organized the reliability effort around the full deployment experience, aligning engineering teams on a shared classification of failures, clear ownership across dependencies, and priorities for corrective work. The teams built dashboards showing where deployments failed and where time was being spent. The goal was to make recurring patterns visible enough to prioritize, rather than requiring another investigation from scratch each time a customer encountered a problem. 

Trustworthy measurement also required separating different kinds of outcomes. A safety check correctly rejecting an unsuitable model was not the same as an infrastructure failure. Gupta distinguished improvements in classification from improvements in the platform itself, so cleaner reporting would not be mistaken for engineering progress. 

He then worked with dependency owners to establish time-based service-level objectives and clear responsibility for each failure domain. Reliability reviews shifted toward prepared accounts of failures, their causes, and the actions needed to prevent recurrence. Each corrective action had an accountable owner and follow-through. “I cannot make an end-to-end commitment to my customer unless my dependencies have measurable commitments to me,” Gupta explains. 

One case illustrated why that discipline mattered. A deployment that had completed its main work remained blocked by an old maintenance operation that had never been marked complete or failed. Closing the stale operation resolved the immediate delay, but the investigation exposed a broader gap: A dependency could remain in a running state indefinitely. Gupta pushed for explicit completion, failure, and timeout conditions for operations that could block a deployment. 

The program also exposed a management challenge. Better diagnosis produced more corrective actions, but those fixes still competed with other engineering priorities. Gupta refined the process to prioritize work by its effect on deployment success, deployment time, and recurring failure patterns, with leads responsible for driving it to completion. Identifying the cause was only useful if the organization made room to address it. 

Over roughly six months, Gupta reports that median deployment time fell from about two days to two and a half hours, while deployment success reached approximately 96 percent. The faster turnaround reduced the time engineers spent waiting for models to reach production.

Reliability is one part of Gupta’s broader AI infrastructure work, which also includes serving efficiency and capacity planning. In a separate initiative he helped lead, teams used a disaggregated serving architecture to improve GPU throughput by approximately 4.4 times for an advertising ranking workload while meeting its latency target. Together, these efforts address practical constraints on deploying AI: Whether a model can reach production predictably and whether the infrastructure can serve it efficiently. 

For Gupta, infrastructure endurance means sustaining reliability as models, workloads, and dependencies change. His deployment program illustrates what that requires in practice: Measurements that reflect the customer’s experience, commitments across team boundaries, and a process that carries a failure investigation through to prevention. Those operating disciplines give engineering improvements a way to become lasting platform capabilities.

This article was written in cooperation with Alexander maxwell