Davide Belli

Staff Research Engineer at Qualcomm AI Research (November 2019 - Present)

I design and implement novel Deep Learning methods across multiple domains including Perception (visual and RF), Model Efficiency, and Agentic AI (LLMs). My work has led to significant innovations:

• Reached 92% of human-data RL performance via self-play without human tasks, using generated tasks with verifiable subtask rewards
• Cut cloud API costs by up to 76% across deep-search, app-control, and embodied tasks, retaining at least 84% of cloud-agent performance
• Introduced structured representations and architectures in AI Agents, reducing model footprint by 30x at the same task accuracy
• Proposed a Dynamic Pruning algorithm for recent LLMs (Phi 3, Llama 3), improving throughput by 40% and memory footprint by 45%
• Implemented Neural Kalman Filter and Weighted Least Squares methods for GNSS localization, reducing positioning error by 50%
• Introduced a personalized biometric neural network classifier that improved accuracy from 90% to 99% AUC
• Integrated solutions in products released to customers (Samsung, Google) and internal departments
• Filed 13 patent applications (3 granted, 10 pending)