Vittorio Rossi
I build LLM systems that run in production and study how generative models spend their compute. I am the founding AI engineer at Commerce Clarity, where I am building the agent development loop end to end: from the infrastructure that now serves dozens of enterprise clients to the agents themselves and the evaluation loops that improve them.
My research is on diffusion language models: how they generate many tokens in parallel, and how to make that cheaper. I tackled their length problem: they must fix the answer's length before decoding, and I showed they already know it after one forward pass, which cuts compute by over 30% with no training. On the side I build IRE, an open-source environment for doing research with AI agents.
I finish my MSc in Artificial Intelligence at Bocconi in December 2026.
Timeline
- 2025 – nowStudent Researcher, Bocconi · with Dirk Hovy · since Jun
- 2024 – nowFounding AI Engineer, Commerce Clarity
- 2024 – 2026MSc Artificial Intelligence, Bocconi
- 2024Exchange, Georgia Tech · Jan – May
- 2023Data Scientist Intern, Q8 Kuwait Petroleum Italy
- 2021 – 2024BSc Economics, Management and Computer Science, Bocconi
News
- Oct 2026Diffusion LMs Are Natively Length-Aware accepted at the NeurIPS 2026 Workshop on Diffusion Language Models.
- Jun 2026Trained ACT robot policies from teleoperation at the EuroTech × Hong Kong hackathon.
- May 2026Released IRE under MIT for macOS and Linux.
Research
Masked diffusion LMs know how long their answer will be before decoding. Reading that from one forward pass and cropping the canvas cuts FLOPs by over 30% on every benchmark, up to 90%, with no training and no significant accuracy loss.
Diffusion samplers reveal several tokens per step and lose the dependence between them. I test whether that makes them prefer more compressed tokenizers than autoregressive models, across training and sampling budgets.
LLMs as zero-shot forecasters against ARIMA and Prophet, with an evaluation protocol for the emergent skill.
Building
A local-first workspace where researchers run Claude Code, Codex or OpenCode behind one interface, with a Git-tracked wiki, structured memory and background experiments.
ACT policies and a LeRobot dataset trained from SO-101 arm teleoperation, published on Hugging Face.
Turns codebases into knowledge graphs so LLMs can retrieve across repositories larger than their context.