Research Intern at Google / Intrinsic · PhD Candidate

I work at the intersection of model optimization, efficient inference, and the mathematics of modern training.

At Google / Intrinsic, I work on model compression, inference optimization, and cross-platform computer-vision deployment for robotic perception. I build reproducible end-to-end workflows spanning PTQ and QAT, distillation, ONNX export, graph surgery, accelerator compilation, runtime integration, and on-device validation across Qualcomm and SiMa.ai accelerator platforms. I profile accuracy, latency, throughput, memory footprint, numerical fidelity, and accelerator coverage to understand the trade-offs required to move research models onto heterogeneous edge hardware.

Alongside this, I am a final-year PhD candidate in Stochastic Optimization for Deep Learning at the University of Basel, supervised by Prof. Dr. Aurelien Lucchi, with expected graduation in October 2026. My research uses stochastic differential equations to understand adaptive optimizers, batch-size scaling laws, compressed and distributed learning, and differential privacy.

Current focus

• Efficient models for robotics: PTQ/QAT, distillation, graph surgery, accelerator compilation, profiling, and cross-platform on-device deployment.
• Optimizer dynamics and scaling laws: connecting SDE-based theory to measurable training behavior and hyperparameter transfer.
• Reliable modern training: adaptive optimization, reinforcement learning, compression, and high-privacy learning.

Before Intrinsic, I built VLM/VLA-based agentic systems for real-world robots at Flexion Robotics, worked on scalable machine learning at Yahoo Research, and spent three years as an AI Quantitative Analyst at UBS.

You can reach me at eneamonziocompagnoni@gmail.com.