Senior AI Applied Scientist · Syllotips · Rome
01 now
Syllotips lets enterprise AI agents learn from a company's own experts. I work on the engine underneath: how agents plan, which tools they call, how they cite their sources, and how we find where they fail. Most of it is the , not the model.
deep research
The main product, redesigned around tool calling, shared working memory and a harness with tool budgets and bounded replanning.
92%judged answer quality,
up from 78%
agent runexample run
tools
Connects the agents to enterprise tools such as Atlassian, HubSpot and Intercom: one architecture for cloud and private cloud, OAuth handling, and tool results the agents can cite.
compute
A sandbox where agents analyse spreadsheets and produce reports and files, on cloud and private cloud alike.
experts
Helps experts close the loop on issues: it investigates what went wrong and proposes a solution for them to review.
serving
vLLM on multi-GPU clusters: open-weight LLMs with up to 1M-token context, plus small decision models for fast routing and grading. The agents can run fully on our own infrastructure.
fine-tuning
Fine-tuned open-weight LLMs from 0.6B to 32B so that one small model replaces three commercial LLM calls for answer confidence estimation.
02 career
Work and study ran in parallel for most of it, like two branches of the same run. Scroll the steps; the strip shows where each one sits in time.
studySapienza · BSc in Computer Science, 107/110
Thesis on knowledge-based Word Sense Disambiguation with synset embeddings.
workCINI · Software Developer
First job: a tool for lexicon-semantic stylometric analysis of large text collections.
workBabelscape · Software Engineer
Built myKnowledgeGraph for knowledge graph extraction, the company's Java NLP pipeline, and concept extraction for trademark names.
studySapienza · MSc in Computer Science, 110/110 with honors
Thesis on language-agnostic latent spaces for sentences, supervised by Roberto Navigli in the Sapienza NLP group.
workBabelscape · NLP Engineer
Built Semantic Paths, multilingual semantic search for scientific papers, and Word Sense Disambiguation models for large-scale multilingual text. Best Employee Award in 2021.
workBabelscape · NLP Researcher
Led Vera, a multilingual fact-checking product on LLMs and RAG. Fine-tuned open LLMs on Slurm clusters for text-to-SQL and claim verification. Worked on Minerva, the first Italian LLM. Mentored interns from prototype to deployment.
studySapienza · Ph.D. in Artificial Intelligence
XXXVII cycle of the National PhD Programme in Artificial Intelligence, as a joint industrial Ph.D. with Babelscape, advised by Roberto Navigli at Sapienza NLP: lexical semantics for multilingual NLP, and word and sentence representations for RAG and LLMs. Defended in Jan 2025.
03 research
Factuality, word senses, and multilingual and multimodal meaning, mostly with the Sapienza NLP group. Published at ACL, EMNLP, NAACL, EACL and in Computational Linguistics. Reviewer for ACL, EMNLP, EACL and ICLR.
Computational Linguistics · 2026 · co-first author
A large-scale resource for end-to-end factuality evaluation, built by extracting and falsifying facts from Wikipedia.
Findings of ACL · 2024 · first author
Finds the ambiguous words in raw text and links each one to its sense, with no pre-marked spans or candidate lists.
EMNLP · 2025
A wide-coverage, semantically annotated multimodal dataset of 165K+ visual concepts built on BabelNet, with the Concept-10k benchmark for vision-language models.
Findings of NAACL · 2025
Part of the Minerva work: adapting the vocabulary cuts token fertility, so the same model reads more Italian per token.
CLiC-it · 2025
What worked when continually pretraining an Italian LLM: data recipes, context extension from 4K to 16K tokens, and instruction tuning.
04 open source
05 writing