포지션 상세
* Who We Are
Video is 90% of the world's data. Most of it is invisible to machines.
TwelveLabs builds the intelligence layer to change that. Our multimodal AI models understand video the way humans do — across sight, sound, and motion — and power production-scale AI workloads across media, entertainment, sports, security, and government.
We have raised more than $210 million from NEA, Radical Ventures, Amazon, NVIDIA, Snowflake, Databricks, Index Ventures, NAVER Ventures, Korea Investment Partners, Quadrille Capital, Red Bull Ventures, and AI pioneers including Fei-Fei Li, Silvio Savarese, and Alexandr Wang.
We are a global company, headquartered in San Francisco with offices in Seoul, New York, and London, and employees around the world. We believe the differences in our cultural, educational, and life experiences make our products stronger. Building technology that understands the world in all its complexity requires people who see it from every angle. We are looking for individuals who are driven by hard problems and want their work to matter. Come build it with us!
* About Jockey
Jockey is TwelveLabs' unified agentic system that reasons across your videos and images. It combines a reasoning model with a memory layer that builds a knowledge store from your corpus.
No context window holds a video archive. We work at a million hours of video. A single model forward pass can tell you about one file; it can't reason across a corpus, and no context window closes that gap. Jockey decomposes a query, retrieves, segments, and reasons across thousands of videos and images. Point it at an archive, ask for a highlight reel or the best viral moments, and it returns timestamped cuts you can use. Corpus-level understanding you can act on is the whole product.
Built for agents, not just people. As AI agents increasingly become the primary consumers of video, we're building production-grade infrastructure that scales to millions of hours while delivering reliable, high-quality results for both human users and autonomous agents.
We build on models we own. Marengo, our embedding model, resolves a query like "the moment we almost missed the flight" into real retrieval. Pegasus, our video-language model, returns structured, timestamped moments on a schema you define. We ship and improve both continuously, so Jockey's quality compounds with every release — no re-integration for customers. Few teams get to build an agent on a stack they control end to end.
Deep expertise, one system, open culture. Foundation models, knowledge construction, search, and the agent harness all live in one org. Each team owns its domain and is expected to have deep expertise in it — but like a Formula 1 team, we optimize for the global system, not local parts. A model gain that doesn't expand what the agent can do isn't a gain. We trace a single algorithm change through to end-system behavior, and share work in progress weekly, not just finished results. Anyone can pull the context they need from any team.
* About the team
The Cognition Models team owns the models that turn video into structured understanding and reasoning: Pegasus, our video-language model, and Jockey Core, the reasoning LLM behind Jockey. In the model stack we sit between Perception Models (embeddings and retrieval) and the agent system — taking what's retrieved and producing structured understanding and the reasoning to act on it.
We focus on multimodal systems with high instruction-following capability and complex, hierarchically structured outputs. Our work spans training infrastructure from pre-training to RL, temporal segmentation and structured metadata extraction, large-scale inference and serving systems, data-curation and evaluation pipelines, and building Jockey Core. We ship products with real-world value rather than doing research in isolation, working as a goal-oriented, cross-functional team of ML researchers and engineers — using the most advanced compute in the world, including NVIDIA B300s, to accelerate the research-to-production cycle.
* About Pegasus
Pegasus is TwelveLabs' video-language model — it turns video into useful analysis by reasoning over visuals, speech, audio, and on-screen text. A key capability is Segment, our time-based metadata feature: instead of a broad question about a video, customers define the exact segment types they care about and the metadata fields they want back, and Pegasus finds the relevant start and end times and returns structured metadata for each segment — titles, summaries, topics, people, visual subjects, confidence, or domain-specific labels. This turns video into time-based, structured data that flows directly into search, archive, editing, compliance, or content-management workflows.
• Design rigorous experiments and evaluation methods that produce clear signals on complex multimodal problems, including where ground truth is ambiguous.
• Strengthen the team's research approach by helping reframe problems, sharpen hypotheses, and raise the bar for experimental rigor.
• Work closely with ML Engineers to translate research advances into production, informing tradeoffs around architecture, serving, and system design.
• Communicate research findings clearly and use them to inform technical direction across the team.
• Explore and adopt AI-assisted development tools such as Claude, Gemini, and GPT to improve productivity across coding, experimentation, debugging, and documentation.
• A track record of driving research on problems with significant technical ambiguity, demonstrated through projects, publications, or technical contributions.
• Strong proficiency in Python and PyTorch.
• Exceptional experimental judgment, including the ability to design evaluations for complex multimodal problems, run rigorous ablations, and draw clear conclusions from empirical results.
• Strong communication skills and a track record of strengthening others' research through collaboration — helping formulate sharper hypotheses, identify more informative experiments, or reframe problems more tractably.
Video is 90% of the world's data. Most of it is invisible to machines.
TwelveLabs builds the intelligence layer to change that. Our multimodal AI models understand video the way humans do — across sight, sound, and motion — and power production-scale AI workloads across media, entertainment, sports, security, and government.
We have raised more than $210 million from NEA, Radical Ventures, Amazon, NVIDIA, Snowflake, Databricks, Index Ventures, NAVER Ventures, Korea Investment Partners, Quadrille Capital, Red Bull Ventures, and AI pioneers including Fei-Fei Li, Silvio Savarese, and Alexandr Wang.
We are a global company, headquartered in San Francisco with offices in Seoul, New York, and London, and employees around the world. We believe the differences in our cultural, educational, and life experiences make our products stronger. Building technology that understands the world in all its complexity requires people who see it from every angle. We are looking for individuals who are driven by hard problems and want their work to matter. Come build it with us!
* About Jockey
Jockey is TwelveLabs' unified agentic system that reasons across your videos and images. It combines a reasoning model with a memory layer that builds a knowledge store from your corpus.
No context window holds a video archive. We work at a million hours of video. A single model forward pass can tell you about one file; it can't reason across a corpus, and no context window closes that gap. Jockey decomposes a query, retrieves, segments, and reasons across thousands of videos and images. Point it at an archive, ask for a highlight reel or the best viral moments, and it returns timestamped cuts you can use. Corpus-level understanding you can act on is the whole product.
Built for agents, not just people. As AI agents increasingly become the primary consumers of video, we're building production-grade infrastructure that scales to millions of hours while delivering reliable, high-quality results for both human users and autonomous agents.
We build on models we own. Marengo, our embedding model, resolves a query like "the moment we almost missed the flight" into real retrieval. Pegasus, our video-language model, returns structured, timestamped moments on a schema you define. We ship and improve both continuously, so Jockey's quality compounds with every release — no re-integration for customers. Few teams get to build an agent on a stack they control end to end.
Deep expertise, one system, open culture. Foundation models, knowledge construction, search, and the agent harness all live in one org. Each team owns its domain and is expected to have deep expertise in it — but like a Formula 1 team, we optimize for the global system, not local parts. A model gain that doesn't expand what the agent can do isn't a gain. We trace a single algorithm change through to end-system behavior, and share work in progress weekly, not just finished results. Anyone can pull the context they need from any team.
* About the team
The Cognition Models team owns the models that turn video into structured understanding and reasoning: Pegasus, our video-language model, and Jockey Core, the reasoning LLM behind Jockey. In the model stack we sit between Perception Models (embeddings and retrieval) and the agent system — taking what's retrieved and producing structured understanding and the reasoning to act on it.
We focus on multimodal systems with high instruction-following capability and complex, hierarchically structured outputs. Our work spans training infrastructure from pre-training to RL, temporal segmentation and structured metadata extraction, large-scale inference and serving systems, data-curation and evaluation pipelines, and building Jockey Core. We ship products with real-world value rather than doing research in isolation, working as a goal-oriented, cross-functional team of ML researchers and engineers — using the most advanced compute in the world, including NVIDIA B300s, to accelerate the research-to-production cycle.
* About Pegasus
Pegasus is TwelveLabs' video-language model — it turns video into useful analysis by reasoning over visuals, speech, audio, and on-screen text. A key capability is Segment, our time-based metadata feature: instead of a broad question about a video, customers define the exact segment types they care about and the metadata fields they want back, and Pegasus finds the relevant start and end times and returns structured metadata for each segment — titles, summaries, topics, people, visual subjects, confidence, or domain-specific labels. This turns video into time-based, structured data that flows directly into search, archive, editing, compliance, or content-management workflows.
주요업무
• Drive research on Pegasus's harder problems such as temporal segmentation, multi-hour context, structured output generation, and training strategies from pre-training through RL, where the right approach requires deep judgment.• Design rigorous experiments and evaluation methods that produce clear signals on complex multimodal problems, including where ground truth is ambiguous.
• Strengthen the team's research approach by helping reframe problems, sharpen hypotheses, and raise the bar for experimental rigor.
• Work closely with ML Engineers to translate research advances into production, informing tradeoffs around architecture, serving, and system design.
• Communicate research findings clearly and use them to inform technical direction across the team.
• Explore and adopt AI-assisted development tools such as Claude, Gemini, and GPT to improve productivity across coding, experimentation, debugging, and documentation.
자격요건
• Significant research experience in one or more areas relevant to video understanding, such as multimodal LLMs, large-scale distributed training, temporal modeling, data-centric model development, computer vision, or vision-language systems, with demonstrated depth in at least one.• A track record of driving research on problems with significant technical ambiguity, demonstrated through projects, publications, or technical contributions.
• Strong proficiency in Python and PyTorch.
• Exceptional experimental judgment, including the ability to design evaluations for complex multimodal problems, run rigorous ablations, and draw clear conclusions from empirical results.
• Strong communication skills and a track record of strengthening others' research through collaboration — helping formulate sharper hypotheses, identify more informative experiments, or reframe problems more tractably.
