Keegan McCallum is the founder of uRun, which builds infrastructure for continuous, interactive generative video. Previously head of machine-learning infrastructure at Luma AI, he helped scale Dream Machine to one million users in four days before turning his attention to making generated video behave like responsive software.
Earlier, McCallum served as head of engineering at Colony Networks and built sql4pandas, an open-source tool that runs SQL queries directly against pandas data frames without copying them into an external database. He subsequently founded Xler.ai, focused on deploying, evaluating, monitoring, and fine-tuning multi-agent applications, and contributed an AutoGen code interpreter using E2B cloud sandboxes to isolate model-generated code execution.
Luma’s Dream Machine launched in June 2024 with approximately 500 H100 GPUs. As demand surged, McCallum and his colleagues expanded to roughly 5,000 H100s within six hours and tapped additional training-cluster capacity. Their original Triton-based architecture struggled with multi-node execution, heterogeneous accelerators, failure handling, and researchers’ development workflows.
McCallum helped develop portable PyTorch-based model serving that separated CPU preparation from GPU execution, enabling distributed workers to pull jobs across providers and accelerator types. Immutable model versions, complete runtime environments, and reversible deployments supported reproducible updates across a large GPU fleet. He also treated scheduling as a product decision: ranking jobs by the percentage of their permitted waiting time prevented lower-priority customers from being indefinitely displaced by enterprise and API traffic. His account of Dream Machine’s scaling challenges connects those architectural choices directly to customer experience.
McCallum subsequently launched uRun with founding engineers Sean Kane and Matt Krzus. Its central technical priorities include:
- Stateful interactive inference: Persistent sessions coordinate globally distributed GPUs, WebRTC connections, user controls, and streaming model pipelines, supporting generated video that responds continuously to its audience.
- Long-horizon video coherence: Small frame-level errors compound during extended generation. McCallum argues that durable visual consistency should be built into models, instead of relying on fragile sampling techniques.
- Generative video as a responsive interface: Interactive avatars, virtual try-ons, camera-driven transformations, and steerable creative tools become practical when models generate faster than people consume their output. Visual interaction may also improve accessibility for users less comfortable with text-heavy interfaces.
- Agent-ready creative infrastructure: uRun combines an embeddable React component, a programmable Python runtime, and CLI and MCP interfaces for assembling real-time visual applications through conventional developer tools or software agents.
His work on real-time generative video focuses on the infrastructure needed to turn increasingly efficient video models into persistent, steerable experiences.
