About the job
You will join the multimodal team to push toward superhuman multimodal intelligence. Advance understanding and generation across modalities—image, video, audio, and text—spanning the full stack: data curation/acquisition, tokenizer training, large-scale pre-training, post-training/alignment, infrastructure/scaling, evaluation, tooling/demos, and end-to-end product experiences. Collaborate cross-functionally with pre-training, post-training, reasoning, data, applied, and product teams to deliver frontier capabilities in multimodal reasoning, world modeling, tool use, agentic behaviors, and interactive human-AI collaboration. Contribute to building models that can see, hear, reason about, and interact with the world in real time at unprecedented levels.
Responsibilities
Design, build, and optimize large-scale distributed systems for multimodal pre-training, post-training, inference, data processing, and tokenization at web/petabyte scale.
Develop high-throughput pipelines for data acquisition, preprocessing, filtering, generation, decoding, loading, crawling, visualization, and management (images, videos, audio + text).
Advance multimodal capabilities including spatial-temporal compression, cross-modal alignment, world modeling, reasoning, emergent abilities, audio/image/video understanding & generation, real-time video processing, and noisy data handling.
Drive data quality and studies: curation (human/synthetic), filtering techniques, analysis, and scalable pipelines to support trillion-parameter models.
Create evaluation frameworks, internal benchmarks, reward models, and metrics that capture real-world usage, failure modes, interactive dynamics, and human-AI synergy.
Innovate on algorithms, modeling approaches, hardware/software/algorithm co-design, and scaling paradigms for state-of-the-art performance.
Build research tooling, user-friendly interfaces, prototypes/demos, full-stack applications, and enable rapid iteration based on feedback.
Work across the stack (pre-training → SFT/RL/post-training) to enable reasoning, tool calling, agentic behaviors, orchestration, and seamless real-time interactions.
Qualifications
Minimum
Hands-on experience with multimodal pre-training, post-training, or fine-tuning (vision, audio, video, or cross-modal).
Expert-level proficiency in Python (core language), with strong experience in at least one of: JAX / PyTorch / XLA.
Proven track record building or optimizing large-scale distributed ML systems (training/inference optimization, GPU utilization, multi-GPU/TPU setups, hardware co-design).
Deep experience designing and running data pipelines at scale: curation, filtering, generation, quality studies, especially for noisy/real-world multimodal data.
Strong fundamentals in evaluation design, benchmarks, reward modeling, or RL techniques (particularly for interactive/agentic behaviors).
Proactive self-starter who thrives in high-intensity environments and is passionate about pushing multimodal AI frontiers.
Willingness to own end-to-end initiatives and do whatever it takes to deliver breakthrough user experiences.
Preferred
Experience leading major improvements in model capabilities through better data, modeling, algorithms, or scaling.
Familiarity with state-of-the-art in multimodal LLMs, scaling laws, tokenizers, compression techniques, reasoning, or agentic systems.
Proficiency in Rust and/or C++ for performance-critical components.
Hands-on work with large-scale orchestration tools such as Spark, Ray, or Kubernetes.
Background building full-stack tooling: performant interfaces, real-time research demos/apps, or end-to-end product ownership.
Passion for end-to-end user experience in interactive, real-time multimodal AI systems.