I am a physicist based in Rome. I hold a Ph.D. in Physics, specifically centered in reservoir computing and the modelling of complex spatio-temporal systems: echo state networks, computational neuroscience, whole-brain mean-field models, and climate forecasting. I'm now transitioning that dynamical-systems mechanistic interpretability framework toward technical AI safety, I'm very interested in understanding how complex, high-dimensional systems fail, drift, and lose predictability, and thus whether the explanations we produce for large language models can be trusted, or whether the agents built on top of them behave the way they claim to.

• AI Safety & interpretability. Faithfulness evaluation of feature attributions for LLMs, with an emphasis on the identifiability problems that arise when perturbation-based explainers and perturbation-based evaluators are confounded by construction — i.e. when an "explanation" and the metric that scores it can't actually be told apart.
• Verifiable and trustworthy agentic systems. Tooling that makes agent behavior auditable and failure-aware rather than opaque: gating actions on machine-checked evidence instead of model self-report, and giving agents an honest, abstaining memory of past failures rather than a confident-but-wrong one.
• Dynamical-systems methods for sequence models. State-space and attractor-geometry perspectives on modern architectures, carried over from my reservoir-computing work — a lens I think is underused in understanding robustness and failure modes of large models.

reservoir-countermanding (Julia).

I developed a Reservoir-Computing (RC) digital twin of macaque dorsal premotor cortex, inferred directly from multi-unit recordings during a stop-signal task. The autonomous model reproduces trial-level neural dynamics and enables virtual experiments linking pre-Go cortical state to reaction-time variability, movement inhibition, and stop-signal reaction time. Our results establish a mechanistic link between latent cortical dynamics and flexible behavioral control, demonstrating how autonomous neural network inference can uncover circuit-level computations.

In search for the invisible: motor inhibition in monkey premotorcortex and its RNN replicas. biorXiv (2025). doi.org/10.1101/2025.11.24.690225

cross-scale-reservoir-computing (Julia).

I co-developed with @Gabriele a julia package for single- and multi-layer echo state networks with spatial blocking, for forecasting high-dimensional spatio-temporal systems such as the Kuramoto–Sivashinsky equation and sea surface temperature fields. Repo includes deep ESN and next-generation RC baselines and a full multi-resolution data pipeline. This is the code behind the peer-reviewed paper below.

Cross-scale reservoir computing for large spatio-temporal forecasting and modeling. Neurocomputing (2026). doi:10.1016/j.neucom.2026.133849

llm-peer-influence-demo (Python - Demo Project).

I use a statistical-mechanics lens — the spin/magnetization formalism from opinion dynamics — to ask whether an LLM agent revising its answer after seeing peers is really responding to how many peers hold each view, or to what those peers said. That distinction matters for multi-agent oversight: debate relies on agents (and judges) tracking argument content, while voting-based aggregation degrades once errors correlate. Holding the vote split fixed and varying only the arguments attached to it, across seven instruction-tuned models and 1.5M+ responses, I find argument content shifts the outcome in roughly 80% of tested cases, beyond what peer count alone predicts, and that this effect is amplified under majority pressure.

never-again (Python). A local-first failure memory for coding agents, exposed over the Model Context Protocol. When an agent hits an error, it can check whether that same failure was already solved in an earlier session instead of re-debugging it from scratch; solved fixes are written back with a compact prevention rule. Runs on a local SQLite file with no server and no data leaving the machine, blends keyword and optional semantic search with reciprocal rank fusion, and — importantly — abstains when the match is weak rather than returning a confident wrong answer.