→ grounded agents
→ meaning & factuality
I build AI agents that stay grounded in production.
At Syllotips I work on the harness that governs how agents plan, act, and stay grounded: orchestration, guardrails, memory and tooling, so an agent reasons over a real company instead of hallucinating its way through it.
Before this, 8 years at Babelscape, from software engineer to NLP researcher, and a Ph.D. in AI with Sapienza’s NLP group on how machines pin down what words actually mean in context. That work landed at ACL 2024 and in Computational Linguistics.
Currently
What I’m building now
Most “AI agents” demo well and fall apart the moment a real workflow touches them. My job is the harness around the model: the unglamorous middle that makes an agent’s reasoning trustworthy, repeatable, and supervisable.
Agent orchestration & governance
Agents that plan multi-step work, recover when a step fails, and stay grounded: they cite what they actually know and say so when they don't, instead of making things up. Rebuilding our deep-research agent this way raised judged answer quality from 78% to 92%.
Agentic context engineering & memory
Shaping behaviour through evolving context and durable memory. What an agent carries between turns is what makes it reliable.
MCP & tool execution
Building Model Context Protocol servers, and running agent-written code in sandboxed environments so agents can read, transform and generate complex files, then call real tools safely with the human still in the loop.
Continuous improvement
Subject-matter experts correct an agent once; those corrections become durable, governed knowledge the whole system learns from. No retraining cycle required.
Small models on our own hardware
Hundreds of fine-tuning runs on a supercomputer, from 0.6B to 32B parameters, so one small model can replace three commercial LLM calls for answer confidence estimation. Served with vLLM on our own GPUs.
Evaluation that decides what ships
LLM-judged comparisons on public and internal benchmarks. A fix that does not move its target cases gets reverted, however good it looked in a demo.
Background
Research & open source
LLM-OASIS · Computational Linguistics 2026
Co-first author on the largest resource for end-to-end factuality evaluation: can a system tell whether generated text is actually faithful to its sources, not just fluent? The grounding problem behind every agent I build now.
Minerva · first Italian LLM
Worked on supervised fine-tuning, vocabulary adaptation and continual training. Adapting the vocabulary cut token fertility for Italian, so the same model reads more text per token.
Word Sense Linking · ACL 2024
Disambiguating word meaning on real, messy text, not just curated benchmarks. The hard part isn’t the dictionary; it’s deciding which sense a sentence actually triggers.
ConceptPedia · EMNLP 2025
A large-scale resource connecting concepts across languages, built to give models a shared semantic backbone instead of a per-language patchwork.
Beanis · Redis ODM for Python
A Pydantic-style typed ODM for Redis: ~70% less boilerplate, performance within 8% of raw Redis. For people who want type safety without giving up speed.
I like the seam between research and production: taking something that works in a paper and making it survive contact with messy data and tight latency budgets. When I’m not on that, I’m writing it up or shipping it as open source.
Recent
News
| Jun 15, 2026 | Building Syllotips’ governed agent runtime: orchestration, Model Context Protocol tooling, and agent memory that keeps reasoning grounded and supervisable. |
|---|---|
| Nov 07, 2025 | ConceptPedia accepted at EMNLP 2025: a large-scale, cross-lingual concept resource for grounding semantic understanding. |
| Oct 23, 2025 | Released Beanis, a typed Redis ODM for Python: ~70% less boilerplate, performance within 8% of raw Redis. |
| Aug 12, 2024 | Word Sense Linking presented at ACL 2024: disambiguating word meaning on real, in-the-wild text. |
Writing
From the blog
Peer-reviewed