Keegan McCallum founded uRun to build infrastructure for interactive generative video. His earlier work on Luma’s Dream Machine connects large-scale model serving with the scheduling and deployment decisions that determine a user’s experience.[3][6]
Start with Generative Video at the Speed of Light →
Education
Links
Owned professional account 
Linked owned GitHub profile Original conference career biography Original uRun session architecture Founder · uRun
Career
McCallum previously led engineering at Colony Networks and machine-learning infrastructure at Luma AI, and founded Xler.ai for deploying and evaluating agent applications. His open-source work includes sql4pandas, which queries pandas data frames through SQL. At uRun, he is applying that infrastructure experience to systems that must respond continuously as a person speaks or changes an input.[3]
Contributions
Keep a streaming interaction in one session
His generative-video talk describes a persistent session joining model execution, user controls and bidirectional streaming. Its Python pipelines connect stages such as microphone input, a realtime voice model, avatar rendering and synchronized audiovisual output. uRun’s session keeps context across the interaction so the application can react while the stream continues, instead of rebuilding a separate job for every input. The current platform offers access by request.[5][7]
Separate preparation, execution and fair scheduling
In his Dream Machine talk, McCallum explains why CPU preparation and GPU execution need different workers. A global queue can distribute work across providers, but preparation must not pull so much into one cluster that it strands jobs behind busy GPUs. A dispatch limit bounds that intermediate backlog. Ranking requests by the fraction of their permitted waiting time also keeps lower-priority traffic from being displaced indefinitely. Versioning the model together with its Python environment makes a deployment reproducible and reversible.[6]
Tools, writing & talks
Writing, research & talks
Start here
Generative Video at the Speed of Lightai.engineerA technical walkthrough of persistent sessions and streaming model pipelines for interactive video.
Trace the voice-to-avatar example from microphone to streamed output and identify where state persists. Familiarity with Python and inference serving helps; the talk explains the architecture before the platform’s access-request step.
Presenter[5]Ten dollars now buys about three hours of continuously generated video, and the argument is that steerability rather than clip quality is the axis that matters, served by a distilled model streamed over WebRTC from globally placed GPUs.
Catalog workOriginal source- Dream Machine: Scaling to 1M Users in 4 Daysai.engineer
A first-person account of distributed model serving, queue fairness and reproducible runtime deployments.
Presenter and former infrastructure lead[6]