Deep Learning Researcher & Engineer
I study how optimization shapes deep learning and make models run efficiently on real hardware.
I am a Research Intern at Google (Intrinsic) in Zürich, working on model optimization and deployment for robotic perception. My work spans PTQ/QAT, distillation, ONNX graph surgery, profiling, and on-device deployment across Qualcomm and SiMa.ai.
I completed my PhD in Computer Science at the University of Basel in November 2026, graduating summa cum laude, under the supervision of Prof. Dr. Aurelien Lucchi. My research uses stochastic differential equations to understand optimizer dynamics, hyperparameter scaling laws, distributed learning, and differential privacy.
Selected work
- No Bias Stereo: optimized stereo inference on dual Qualcomm NPUs, reaching 10.76 FPS (3.75× baseline throughput) through graph surgery, operator fusion, and selective quantization.
- 3PTv2: reduced multi-view 6D pose-estimation latency on NVIDIA A100 from 6.8 to 3.1 seconds per frame (54%); developed a separate Qualcomm portability proof of concept.
- Flexion Reflect v0: built agent components and integrated sim-to-real workflows, contributing to an autonomous-operation demonstration on a physical robot.
Read about my contributions, methods, and results →
Research
My work connects mathematical models of optimization with experiments on language and vision models. I have seven first-author conference papers across ICML, ICLR, AISTATS, and IJCNN, including an AISTATS 2025 Oral.
- Optimizer dynamics and scaling laws: understanding how noise and adaptivity shape training, and how hyperparameters should change with batch size.
- Distributed learning: studying the interaction between communication compression, gradient noise, and stability.
- Private and robust optimization: understanding how privacy noise affects optimization and hyperparameter transfer.
Publications and research details →
Selected research talks
- Google DeepMind, 11 August 2026 — Understanding Optimizer Dynamics and Scaling Laws via SDEs.
- Caltech, Anima Anandkumar’s Lab, 5 March 2025 — Unbiased and Sign Compression in Distributed Learning: Comparing Noise Resilience via SDEs.
- Nokia Bell Labs, 29 January 2025 — Unbiased and Sign Compression in Distributed Learning: Comparing Noise Resilience via SDEs.
All talks and conference presentations →
Background and contact
Before Google (Intrinsic), I worked on robot autonomy at Flexion Robotics, scalable machine learning at Yahoo Research, and applied machine learning in finance during three years as an AI Quantitative Analyst at UBS. My background spans mathematics and quantitative finance; I have also taught deep learning and optimization and mentored master’s research students.
You can reach me at eneamonziocompagnoni@gmail.com.
