Vlad Tsvang

Vlad Tsvang has been a frontend engineer at Meta since 2020. He helped build an AI teammate that stays with an on-call rotation when engineers change shifts. An incoming engineer can pick up an investigation in progress, see where a proposed fix came from, and correct advice that the next person on call might otherwise follow.

Vlad’s earlier web work included interactive graphics for Euro 2012 and the 2012 Olympics, a joint project for Sports.ru and Gillette. He also built refilterable, an open-source React library for pages that store filter settings in the URL. A developer can define how to read and validate a filter once, then reuse it in several controls. Each component subscribes to the parameters it uses, so changing an unrelated filter leaves that component alone.

Sharing evidence during an incident

At Meta, Vlad contributed to DrP, a platform that lets engineers turn a diagnostic procedure into code. When an alert fires, DrP can run the procedure and attach its findings to the alert, giving the responder a place to start.

With Chinmay Gandhi and Patrick Walsh, Vlad helped build the Agent Feed for investigations in which several specialist agents each have part of the evidence. Through a shared API, agents can publish findings and read what others have found. A coordinating agent ranks competing explanations by checking which ones independent sources corroborate. Responders see the best-supported findings first, and their feedback helps determine which findings the feed promotes.

What the next shift inherits

Vlad, Swaapnika Motapothula, and Chinmay Gandhi worked together on Meta’s AI on-call teammate. Vlad argued for giving each rotation its own agent identity and memory. The team decides what the agent can access, and any engineer taking a shift can use the team’s playbook and the decisions earlier responders recorded.

In Vlad’s example, one engineer teaches the agent to roll back a configuration change first if checkout starts timing out after that change. When the same alert returns days later, the agent can show the next engineer where the advice came from and draft the rollback for review. If the advice proves wrong, the engineer’s correction reopens the investigation and becomes part of the shared memory. The system also scores retrieved entries by whether they helped, promoting useful guidance and retiring stale advice.