What I do
I’m Carlo Nicolini, an AI research scientist and statistical physicist. I help teams with two kinds of hard problem:
- Reliable AI reasoning — interpretability, inference-time methods, evaluation, and probabilistic language programming for LLM and compound AI systems that need to be auditable and predictable, not just impressive on a demo.
- Quantitative portfolio optimization — I maintain skfolio, the Python library for modern portfolio allocation, and build risk models and backtesting that hold up under real constraints.
What ties them together is a habit from physics: turning a fragmented technical space into clear abstractions, operators with guarantees, and software others can rely on. If you’re working on something in these areas, I’m always happy to talk about projects and collaborations — the best way to reach me is on LinkedIn.
About me
I am a senior AI research scientist working at the intersection of artificial intelligence, statistical physics, computational neuroscience, and information theory. For over 15 years, my work has focused on problems that require both theoretical clarity and reliable execution: building models, abstractions, and scientific software that help researchers and engineers reason about complex systems with more rigor. Over time, this has evolved into a form of technical leadership grounded in framing hard problems well, creating reusable foundations, and helping collaborative work move with more clarity.
A recurring theme in my work is taking a fragmented technical space, identifying the right abstractions, and turning them into tools, APIs, and explanations that make an entire line of work more usable.
Scientifically I am interested in modern artificial intelligence from the point of view of a statistical physicist (I started with the great McKay book). I have been working on LLM systems, vector symbolic architectures, mechanistic interpretability, AI reliability, probabilistic language programming, and the design of compound systems that make inference-time behavior more auditable, controllable, and semantically grounded.
In parallel, I have developed quantitative finance models, implemented in the libraries scikit-portfolio first, then becoming a maintainer of skfolio, a large collaborative project to improve the scientific foundations of portfolio allocation with modern optimization methods.
In my recent writing, I explore how branching, verification, and decomposition shape the reliability of AI systems, how inference-time scaffolds can be understood as structured probabilistic procedures, and how better interfaces between theory and engineering can make advanced systems more legible and robust.
My research portfolio extends to online convex optimization, information theory and large scale machine learning models. Across these projects, I tend to work at the boundary between research, engineering, and real-world constraints: framing the problem clearly, choosing the right level of abstraction, aligning technical choices with practical goals, and turning ideas into systems that are mathematically grounded, computationally efficient, and genuinely useful in practice.
My contributions
I have been working across research, engineering, and open-source software, with publications spanning modern AI, machine learning for finance, and complex systems. My work appears at venues such as COLM, ICAIF, TMLR, EPJ Data Science, Physical Review E, and NeuroImage, and is collected on Google Scholar.
- Senior AI Research Scientist, Ipazia SpA (2022–present)
- Maintainer of skfolio (portfolio optimization in Python)
- Research: reliable AI systems, interpretability, compound AI, NLP, optimization, complex systems
- Focus: connecting deep theory, reusable software, and collaborative research execution
- Strengths: technical direction, cross-functional collaboration, and building research infrastructure others can rely on
Selected publications
Recent work on LLM interpretability, NLP, and statistical physics. Full list →
- Sokobench: Evaluating Long-Horizon Planning and Reasoning in Large Language Models
- Unveiling LLMs: The evolution of latent representations in a dynamic knowledge graph
- Hopfield Networks for Asset Allocation
- Glitter or gold? Deriving structured insights from sustainability reports via large language models
Latest blog posts
I maintain a research blog where I develop ideas in public, document ongoing work, and refine the conceptual foundations behind the systems and tools I build.
-
La mente che conta fino a dieci
-
A controlled testbed for Repeat-Your-Self — SAT solving, message passing, and a predictive rho/phi theory
-
Deep Equilibrium Models as a model for RYS
-
Fractal Connectivity Scaffold Networks
-
Hippocampal scaffold networks
Contact
If you are working on reliable AI systems, mechanistic interpretability, inference-time methods, or quantitative portfolio optimization, I’d be glad to hear about it.
The best way to reach me is on LinkedIn.