Computational linguistics · Made in Italy

AI and NLP
for the Italian language.

We build natural language processing for Italian, on our own lexicon and algorithms. We tag every word, find the entities, detect polarity and register, and anonymise documents before they are shared or published.

NounsVerbsAdjectivesArticles & prepositionsAdverbs
italica-X · sample output

Speed · measured

A sentence analysed in 2 milliseconds.

italica-X was designed for Italian and to process large volumes of text quickly. These are the times the engine recorded in 2026, across 1.5 million real requests.

One sentence
2ms

average time for a full analysis: words, lemmas, parts of speech

A document of a few pages
0.6s

average time for a text of 1,000 to 5,000 words

One document anonymised by ADA
0.46s

~50,000 documents analysed in 6.2 hours

The lexicon
4.5million

inflected forms, generated from 115,924 lemmas

Averages calculated from the italica-X request log, January–September 2026. Times measure processing inside the engine, excluding the network. The ADA figure comes from the benchmark of 26 August 2026 on a domain corpus, one document at a time.

One engine, two products

italica-X is the technological core of our applications. italica puts linguistic analysis in a learner's hands, and ADA puts it to work inside public‑sector document systems.

For learners and teachers

italica

Grammatical analysis of Italian in a chat interface. Type a sentence and every word is tagged with its part of speech, lemma and features. The engine takes about 2 milliseconds per sentence, so answers are instant. The interface is available in seven languages. Use it on the web, on iOS or on Android.

Open italica.app →
For public sector, health and justice

ADA

Anonymises your Italian documents locally: 21 entity classes detected, codes validated in real time with deterministic results.

Try ADA on a document →
powered by
The engine

italica-X

A proprietary vocabulary and a set of algorithms built for Italian grammar. It works out each word's grammatical role from its context, so in la porta è chiusa (the door is closed) and porta la borsa (carry the bag) the same word gets two different analyses. italica-X can of course be used to analyse and classify email and PEC (Italian certified email).

API access for integrators →
Tokenization & POStoken → lemma → part of speech
Entity extractionpeople, bodies, places, dates
Polaritypositive · neutral · negative
Registerformal, bureaucratic, colloquial

ADA · On-premises anonymisation

Every document processed and anonymised on premises.

Upload a PDF or a text. ADA analyses it with italica-X, recognises the personal data and produces a version where it is replaced by -OMISSIS-. Each stage stays a separate layer of the document (uploaded, tokenised, annotated, anonymised), so the work can be checked and repeated.

Why choose ADA?

  • SpeedUnder half a second per document: 48,464 documents anonymised in 6.2 hours, one at a time.
  • Data stays putInstalled in your own data centre on Linux machines: CPU only, no GPU, no outbound connection. It's a setup that also suits health data (GDPR art. 9) and criminal-record data (GDPR art. 10).
  • AuditableThe same document always gives the same result. Every detection traces back to explicit rules that can be inspected and checked.
  • Human reviewReplacing data with -OMISSIS- doesn't always make a document anonymous: the context can still point to someone. That's why every annotation can be checked and corrected by hand in the editor before the document is used: in sensitive cases, a person has the final say.
  • Validated codesTax codes (including omocodici), IBANs, number plates, postcodes, ID cards, IP and MAC addresses are detected and then checked against the formal rules of each format.
  • ManagementMulti-tenant, user roles, case files, logs, sharing, export and much more.
Documento di esempioUfficio anagrafe

Richiesta di certificato di residenza

Il sottoscritto Mario Rossi, nato a Roma il 12/04/1978, codice fiscale RSSMRA78D12H501M, residente in via Roma 14, chiede il rilascio del certificato di residenza per uso amministrativo.

PersonaLuogoDataCodice personale · validatoIndirizzo
Personal data highlighted

ADA and local LLMs, side by side

We compared ADA with three local language models, from 7.5 to 27.8 billion parameters, on 50 synthetic legal documents, on the same computer and offline. The video sums up the results: data left readable, speed, energy use and cost.

ADA and three local LLMs on the same 50 synthetic documents
SystemSensitive data left readablePages per minuteSame result across runsPeak memoryEnergy per 1,000 pages
ADAdedicated engine0.3%292100%1.1 GB0.7 Wh
gemma47.5B parameters5.5%3.9100%11.2 GB178 Wh
qwen3.59.7B parameters2.9%3.392%18.0 GB215 Wh
qwen3.827.8B parameters0.7%0.992%27.1 GB542 Wh

Benchmark on an Apple M4 Max with 128 GB of memory, all local; 3 runs per system, LLMs at temperature 0 with a fixed seed. Sensitive data left readable: share of personal data items with at least one part still readable after anonymisation. Pages of 1,800 characters; net energy measured at the power adapter. Videos are 3 minutes, about 9 MB each. Fictitious documents; methodology and data available on request.

ADA for publishing

Legal publishers, databases and newsrooms publish rulings, records and documents that contain personal data every day. ADA anonymises them before publication: names, tax codes, addresses and other personal data become -OMISSIS-, while the rest of the text stays intact, readable and quotable.

  • Whole archives48,464 documents anonymised in 6.2 hours: an archive that size is processed overnight, not over months of manual work.
  • Same rulesThe same data item is masked the same way across the whole archive, document after document.
  • ReviewEditors check the annotations in the editor and correct them before publishing.
  • Confidential textsManuscripts and unpublished documents stay inside your network.

21 entity classes

Identity
person, minor, address, identity document, personal code
Contact and digital
contact, digital data, internet entity
Sensitive (art. 9)
health data, judicial data, family data
Economic and employment
economic data, employment data, VAT number
Procedural
protocol number, proceeding, case file, legal reference
Contextual
organisation, place, date

In development: reversible pseudonymisation, with typed placeholders (PERSONA_001) and an encrypted lookup table kept separately.

Appliance · italica-X and ADA

Everything on one machine, in your own data centre.

italica-X and ADA come preinstalled and ready to use. You can have them on a physical server or as a virtual machine to import into your own infrastructure. The software is the same either way.

Physical server

Delivered configured

Arrives with the software installed; it only needs to join your network. Suits organisations without a virtualisation platform.

Virtual machine

Ready-made image

An image to import into your hypervisor, for organisations that already run a data centre or private cloud.

What's inside, what it needs

  • Softwareitalica-X, the language engine with a lexicon of 4.5 million inflected forms, and ADA, the anonymisation web application.
  • HardwareLinux, CPU only. No GPU and no accelerator.
  • NetworkNo outbound connection. No external service in the processing chain.
  • AccessADA from a browser on the internal network. italica-X through an HTTP API, to build the analysis into your systems.
  • Speed0.46 seconds per anonymised document, measured on a domain corpus.

Sizing and configuration: get in touch →

On-premises LLM · Design

Language models installed in your own infrastructure.

We design solutions built on large language models that run on the client's own infrastructure. Documents, questions and answers never leave your network. For public administrations, healthcare and companies handling confidential data, that's a solid basis for data protection, together with the organisational measures and the legal basis for the processing.

  1. AnalysisUse cases, volumes, security constraints and the hardware available.
  2. Model choiceOpen-weight models evaluated on Italian and on your documents, not on generic benchmarks.
  3. ArchitectureSearch over internal documents (RAG), access control, logging and audit, in your data centre or private cloud.
  4. RolloutInstallation, quality checks in Italian, staff training and maintenance.
Your infrastructure
Documents and requestsprotocol, archives, service desks
italica-Xentities extracted, personal data masked
Open-weight LLMon dedicated GPUs, with RAG over internal documents
Answers and draftswith cited sources and an audit log

No calls to external APIs

Research and development

One language, studied in depth.

Our research starts from Italian grammar, not from a general-purpose model. We build the lexicon, write the rules and measure the results on real texts: that's where every feature of italica-X and ADA comes from.

  • Lexicon first115,924 lemmas generate 4,545,631 inflected forms, 39 per lemma on average. Every analysis builds on it.
  • Explicit, verifiable and traceable rulesEvery detection traces back to a rule you can read and check. Analysing a text always produces the same result, on every run.
  • Measurements, not promisesWe only publish measured times: 2 ms per sentence across 1.5 million real requests, 0.46 s per anonymised document.
Example · from lemma to forms

anonimizzare

anonimizzoanonimizzianonimizzaanonimizziamoanonimizzateanonimizzanoanonimizzavoanonimizzaianonimizzòanonimizzeròanonimizzereianonimizzassianonimizzandoanonimizzatoanonimizzataanonimizzati…

115,924 lemmas → 4,545,631 forms

Development status of our technologies

  • In productionGrammatical analysis, entities, polarity and register with italica-X.
  • In productionAnalysis and classification of email and PEC.
  • In productionAnonymisation with ADA across 21 entity classes.
  • In developmentReversible pseudonymisation: typed placeholders (PERSONA_001) and an encrypted lookup table.
  • Per projectOn-premises open-weight LLMs, with italica-X upstream to mask personal data.

SAC3 · Framework agreement for local government

intelligo in the SAC3 framework agreement for public administration.

intelligo is a member of the temporary grouping of companies (RTI) led by IBM Italia that won Lot 1 of the Framework Agreement for Cloud Application Services, edition 3, for Italian local public administrations.

Within the grouping, intelligo contributes its NLP expertise for Italian: text analysis with italica-X, document anonymisation with ADA, and solutions built on large language models that run on the authority's own infrastructure, so the data never leaves it. An authority activates the services through an implementing contract, based on its own needs plan.

SAC3 with IBM

  • IBM Italia S.p.A.Lead
  • intelligo S.r.l.Member

The 13 SAC3 catalogue services

Core

  • SVISviluppo e manutenzione evolutiva
  • MIMigrazione applicativa al cloud
  • CFConfigurazione di soluzioni di terze parti
  • CWSviluppo in co-working con l'ente
  • MADManutenzione adeguativa
  • MACManutenzione correttiva
  • SSSupporto tecnico-specialistico ICT
  • GAGestione del portafoglio applicativo

Add-on

  • GOGestione operativa
  • IAMGestione dell'identità e dell'accesso
  • ACDAcquisizione e classificazione dati
  • ELEE-learning e assistenza virtuale
  • HDContact center e help desk

Data acquisition and classification also covers unstructured data, which is where italica-X and ADA work.

The company

Scientific expertise, proprietary technology.

Founded in 2023 by Antonio Guerrieri, intelligo is a team of engineers who have worked on NLP in universities and in industry. We focus on one language: Italian.

  • FORUM PA 2025

    In Rome, with N&C, as guest partners at the Regione Puglia stand, we presented italica-X and ADA. This year's theme: “Verso una PA aumentata” (towards an augmented public administration).

  • Salento Biomedical District

    Presenting italica and ADA to HealthTech investors in Lecce.

  • Puglia Women AI

    Gold partner of the first women's AI bootcamp, at the Fiera del Levante in Bari.

Founded
2023
Based in
Veglie, Lecce
Quality
ISO 9001
Gender equality
UNI/PdR 125
Language
Italian, first
Status
Startup innovativa

Contact

Bring us a document. We’ll show you what italica-X can do.