I'm Mirko Tritella, researcher and engineer. I work on AI interfaces to official records and open government data, designed so that every answer can be traced back to its source.
Why this work
Parliaments, ministries and city councils publish enormous digital records: debates, votes, legislation, budgets, open datasets. This material is public in principle but hard to use in practice. It is scattered, verbose and disconnected, and most of it is never read.
AI can become a new way in. It lets anyone ask a question in plain language and get an answer drawn from thousands of documents. It also creates a new problem: the answer is easy to obtain and hard to check. A system that summarizes the public record without showing its sources replaces one opacity with another.
So I build systems where generation is the last step, not the whole product. Structured data and knowledge graphs keep track of who said what, in which role; retrieval finds the evidence; and every claim is checked against the official record before it reaches the reader.
Democracy depends not only on access to information, but on the ability to understand and question it.
As institutions produce ever larger digital records, AI can help citizens navigate them, provided its answers remain transparent, contextualized and verifiable. That is the space this work lives in.
Featured project
An AI interface for exploring the Italian Parliament through its official records.
The Italian Chamber of Deputies publishes a stenographic report of every plenary session, and every roll-call vote, as open data. Together they can answer almost any question about what parliament actually does. In practice, finding who said what, in which debate and in which role means reading hundreds of pages.
ParliamentRAG lets you ask in plain language. It retrieves the relevant speeches and votes, writes an answer that covers every parliamentary group, majority and opposition alike, and links each quotation back to the exact passage in the official record.
The answer should never be detached from the record.
In a blind evaluation against Google NotebookLM, judged by six domain experts, the system scored higher on parliamentary-group coverage and citation faithfulness, with human ratings favouring it on source and balance dimensions. The methodology and full results are in the research paper.
705 plenary sessions · 174k+ speech chunks · 17.3k roll-call votes · 6.9M individual vote records
XIX Legislature, updated continuously from official open data
The work is fully open: the knowledge graph is published as an RDF dataset on Zenodo (CC BY-SA 4.0) and in tabular form on Hugging Face, with project terms under the w3id.org/parliamentrag namespace. The system is also an MCP server, so AI assistants can query the official records directly.
Where it started
My interest in public data started locally. In Roseto degli Abruzzi, my home town, election results existed only as PDFs on the municipal website; I extracted them and republished them as machine-readable data. During the pandemic I built a site that tracked the town's COVID-19 numbers day by day.
ParliamentRAG asks the same question at national scale: how can public information become easier to access and use? First publishing the data, then building the interfaces to understand it.
Results of the 2016 municipal and 2019 regional elections, extracted from PDFs on the official website and republished as machine-readable CSV and JSON.
A site that tracked up-to-date COVID-19 statistics for Roseto degli Abruzzi.
An AI interface to the Italian Chamber of Deputies, with every answer linked back to the official record.
Research
Retrieval-augmented generation grounds generated text in retrieved sources; citation verification then checks each quotation against the record. Grounding without verification is a promise, not a guarantee.
Representing speakers, roles, debates, votes and acts as entities and relationships makes it possible to answer questions that plain text search cannot, and to keep the context around every retrieved passage.
In institutional records, who speaks matters as much as what is said. Authority-aware retrieval models the speaker's role, expertise and mandate instead of treating all text as equal.
A single accuracy number hides what matters: coverage of viewpoints, balance across groups, faithfulness of quotations. Evaluation of these systems is a research problem in its own right.
Retrieval-augmented generation · knowledge graphs · information retrieval · semantic web and linked open data · evaluation of LLM systems.
Currently exploring semantic axes for political-position mapping, and how parliamentary stances evolve over time.
Publications
Can an AI system summarize parliamentary debate while preserving the diversity of voices and keeping every quotation faithful to the official record?
ISWC 2026, In-Use Track · Springer · to appear
The system behind parliamentrag.it: a live AI interface to the Chamber of Deputies where each answer links back to the stenographic record.
ISWC 2026, Posters & Demos · CEUR-WS · to appear
About
I grew up in Roseto degli Abruzzi, on the Adriatic coast. My first projects were about my home town: republishing its election results in machine-readable formats, tracking its COVID-19 numbers day by day. The instinct was the same then as it is now: public information should be usable by the public.
That instinct scaled up. I studied Computer Science at the University of Bologna and Data Science at the University of Milano-Bicocca, where I carried out the ParliamentRAG research, presented at ISWC 2026. What holds the work together is a simple requirement: an answer about public life should be checkable against the public record.
I live between Rome, Milan and Roseto. I collect technical and research notes in Notes.