Keunwoo Choi

Senior Principal ML Scientist

Leading development of large Language Models.

Adjunct Professor

DM-GY 9103/Q Deep Learning for Media

AI Advisor (Former AI Scientist/Director)

Leading an AI music/audio team covering music source separation, audio generation, and 3D audio.

Senior Research Scientist

Large-scale deep learning systems for music and video content creation, social media, and music streaming service. Music research roadmapping for scalable and sustainable growth.

Research Scientist

Music content understanding and discovery for 300M+ users of the global leader in music industry

Researcher

3D audio reproduction systems for immersive acoustic experience

From First Hire to Team Leader

I joined Genentech as the inaugural member of the LLM team, tasked with navigating the rapidly evolving landscape following ChatGPT's initial release. I actually enjoyed the uncertainty—there's something energizing about figuring out real-world constraints and goals to maximize core value. What began as an exploration in uncharted territory has evolved into leading a comprehensive LLM initiative that spans foundational research, model development, and strategic implementation across Roche.

Foundation and Discovery

The early days were marked by profound uncertainty. Investment levels, data acquisition strategies, and training methodologies were all question marks in a field moving at breakneck speed. I identified time as the most constraining factor and adopted a learning-by-doing approach, recognizing that LLMs are inherently empirical and implicit systems.

Delivered: Complete training pipeline, data frameworks, resource management, initial model prototypes.

Earned: Deep understanding of large-scale model development, hands-on experience with large-scale training, and strategic thinking around resource constraints in rapidly evolving fields.

Scaling and Leadership

The focus shifted to achieving something real—training and deploying LLMs at enterprise scale. I led critical decisions around model selection, defining desired properties, and ensuring comprehensive data coverage. Simultaneously, I transitioned into a management role while navigating the challenge of finding exceptional talent who would not only join our team but thrive and deliver in our demanding, fast-paced environment.

Delivered: Company-wide LLM objectives, evolving hiring processes, team expansion.

Earned: Leadership experience in high-uncertainty environments, advanced hiring and talent identification skills, team management capabilities, and expertise in scaling technical teams during competitive market conditions.

Strategy and Framework

With a larger team in place, my responsibilities expanded to encompass strategic roadmapping for what has become a critical component of the company's AI initiatives. I establish annual goals and quarterly OKRs, supervise team members across all aspects of LLM development, and coordinate with stakeholders throughout the organization.

Delivered: Strategic roadmaps, OKR frameworks, team ownership structures.

Earned: Strategic planning expertise, cross-departmental project management skills, open-source contribution experience, advanced interviewing techniques, and leadership frameworks for distributed technical teams across continents.

Research and Innovation: TalkPlay

Bridging my expertise in music AI with LLM development, I initiated TalkPlay—among the first LLM-native music recommendation systems. This work explores novel research directions in conversational music discovery, pioneering generative recommendation for music through natural language understanding and multimodal token generation.

The research progression began with foundational work that earned us ISMIR 2023 best paper runner-up recognition. In 2024, we published research on dataset generation for conversational LLM music discovery and recommendation at ISMIR 2024. Building on this foundation, we released the first comprehensive TalkPlay paper in 2025, demonstrating how music recommendation can be reformulated as LLM token generation across multiple modalities.

Queen Mary University of London

New York University

Seoul National University

Seoul National University