⚙️Research
I work on AI for design automation. Generative models have become very good at producing artifacts that look like designs. Engineering needs artifacts that are designs: buildable, checkable against hard constraints, verifiable. My research is about closing that gap.
Two mechanisms do the work. Cheap proposals — a language model proposes candidate designs and an optimizer selects among them, so search is no longer limited by human iteration. Cheap evaluation — learned surrogates stand in for the simulators that usually bottleneck the loop. Both are subject to one constraint that makes the difference between a demo and a tool: the output has to be checkable.
I am a PhD student at UC Berkeley in BAIR, advised by Alberto Sangiovanni-Vincentelli and Pierluigi Nuzzo, and I do research on AI for hardware design at Apple.
In Fall 2026 I am a Teaching Assistant for EECS 182/282A: Deep Neural Networks at UC Berkeley. Since 2022 I have taught over 1,000 students across four Berkeley courses.
Reach out in Italian 🇮🇹, English 🇺🇸 or French 🇫🇷.
Education
PhD Computer Science, expected Nov 2027
UC Berkeley — Berkeley AI Research
MSc Electrical Engineering & Computer Science, 2021
Politecnico di Torino · EURECOM · Télécom Paris
BSc Electronic Engineering, 2018
Politecnico di Torino
Focus Areas
What I Work On
AI for Design Automation
Three threads, one question: how do you get a learned system to produce a design that survives contact with a specification?
LLM-Guided Search
A language model proposes candidate designs and an optimizer selects among them, so search is no longer limited by human iteration. EvoScale restructured this loop around the inference cache, raising the prefix cache hit rate from 18.4% to 94.2% and cutting per-candidate generation time 4.7–6.2× on a single H100. In colmo, pairing LLM program synthesis with gradient-free optimization reached target control performance in over an order of magnitude fewer search steps.
Learned Surrogates
Design is search, and search is bottlenecked by simulation. I replace the simulator with something learned: neural operator and spectral surrogates for PDE simulation at Lawrence Berkeley National Lab, and a single-step reinforcement learning framework for analog circuit sizing that cuts runtime more than 8× against DDPG and more than 2× against Bayesian optimization.
Verifiable by Construction
A design you cannot check is not a design. I work on neuro-symbolic methods that satisfy hard constraints by construction rather than by post-hoc filtering, and on measuring trustworthiness against system-level specifications — distilling a driving model to 90% smaller at the cost of 7% more safety violations, and class-wise thresholding that improved out-of-distribution detection robustness to label shift by 20% TPR.
303
Citations
h-index 3 · Google Scholar, August 2026
$130K
Competitive research funding
Berkeley DeepDrive · Meta AI · Fondazione CEUR
1,000+
Students taught at UC Berkeley
Four courses since 2022
6.2×
Faster LLM-guided search
EvoScale, on a single H100
Full list on Google Scholar.
Fall 2026 — Teaching Assistant, EECS 182/282A: Deep Neural Networks. Berkeley’s deep learning course, taught by the EECS department to both undergraduate and graduate students.
Since 2022 I have taught over 1,000 students across four UC Berkeley courses:
| Course | Students |
|---|---|
| EECS 16B — Designing Information Devices and Systems II | 400+ |
| CS 282A — Deep Neural Networks | 300+ |
| EECS 227A — Optimization Models in Engineering | 300+ |
| EECS 249A — Embedded Systems | 70+ |
Before Berkeley I was the first MATLAB Student Ambassador in Italy (MathWorks, 2017–2019), where I built a community of over 1,500 students and ran seminars on control systems, deep learning and linear algebra. Ten further Italian ambassadors were appointed after the role proved successful. I also founded DroneLAB at Politecnico di Torino, raising €3k in sponsorship and leading 15 undergraduates to build four autonomous quadcopters.
Experience
Graduate Student Researcher
UC Berkeley — Berkeley AI Research (BAIR)Neuro-symbolic methods for design automation, advised by Alberto Sangiovanni-Vincentelli and Pierluigi Nuzzo. Work on LLM-guided search for analog circuit sizing and control policy synthesis, and on learned surrogates that replace expensive simulation inside design loops.Research Intern — AI for Hardware Design
AppleDesign automation of hardware using AI, with the Cellular RF and AMS teams. Work under NDA.Research Intern — LLM Alignment
AppleLLM alignment research with the ASE AI/ML team. Work under NDA; contributed to a public preprint on retrieval-augmented query auto-completion.Graduate Student Researcher
Lawrence Berkeley National LaboratoryNeural operator and spectral-method surrogates for PDE simulation, learning resolution-invariant representations of physical systems.Fellow Researcher
EURECOMMedical image segmentation with normalizing flows, with Maria A. Zuluaga. Won a $15k Young Researcher Grant from Fondazione CEUR.
Education
PhD Computer Science
UC Berkeley — Berkeley AI ResearchAdvised by Alberto Sangiovanni-Vincentelli and Pierluigi Nuzzo. Research on neuro-symbolic methods for design automation: LLM-guided search, learned surrogates, and verifiable-by-construction design. GPA 3.97.MSc Electrical Engineering & Computer Science
Politecnico di Torino · EURECOM · Télécom ParisDouble-degree programme. Graduated 110/110 cum laude, top 2% of the cohort. Thesis on out-of-distribution detection in supervised image classification.BSc Electronic Engineering
Politecnico di TorinoSelected for the Young Talent Project (top 5%, best 200 students) for three consecutive years.
Competitive grants, fellowships and merit awards.