Lead Scientist - Real-Time Multimodal Conversational AI

Amazon.com Services LLC; Seattle (WA); On-site; USD 210,000 - 310,000; Full time

Job summary

Amazon.com Services LLC seeks a Principal Applied Scientist to lead real-time multimodal conversational AI research, spanning foundation models for speech and audio and post-training systems that shape natural, human-like behavior.

You will set technical direction, guide architecture for production deployment from inception, and work across pre-training to real-time inference with a focus on low latency and scalable training.

Qualifications

• PhD in Electrical Engineering, Computer Science, Mathematics or related field.
• 10+ years in industrial or academic work building speech recognition and NLP systems.
• Experience training large-scale foundation models with hands-on scaling.
• Experience with multimodal models across speech, text, and audio.
• Strong transformer architecture knowledge in speech/audio domains.
• Publication record at top-tier venues (NeurIPS, ICML, ICLR, Interspeech, ICASSP, ACL).
• Demonstrated leadership: mentoring researchers and guiding teams.

Responsibilities

• Lead Foundation Model Scaling: build/train large multimodal models for real-time speech and audio generation.
• Develop post-training methods: reward modeling and RL for real-time conversational behavior.
• Advance Real-Time Perception & Generation within strict latency budgets.
• Define technical direction and ensure production-ready model deployment from inception.

Skills

PhD in EE/CS/Math

Foundation models

Multimodal models

Transformer architectures

Publications in top venues

Scientific leadership

10+ years in speech/NLP

Education

PhD in Electrical Engineering, Computer Science, Mathematics

Job description

Amazon.com Services LLC seeks a Principal Applied Scientist to lead real-time multimodal conversational AI research, spanning foundation models for speech and audio and post-training systems that shape natural, human-like behavior.

You will set technical direction, guide architecture for production deployment from inception, and work across pre-training to real-time inference with a focus on low latency and scalable training.