-500px.png){kind=spacer}
AI Research
The AI Research Behind ZETTA AI Box
For more than 10 years, ZETTA has built its own AI engines for speech, language, documents and vision. ZETTA AI Box brings that research together in one on-premise platform, with an AI Agent on top.
Research areas
Six areas, one on-premise platform
Speech & Language
Speech recognition, speaker diarization and language identification for real-world, multilingual audio.
Machine Translation
Offline translation across 100 languages, preserving document formatting.
Document Intelligence
OCR and handwriting recognition for complex scanned documents.
Vision
Object detection, key-frame extraction and scene description for images and video.
Agentic AI on the edge
Running large language models and agents locally, on compact hardware, with no cloud dependencies.
Model adaptation
ZETTA Factory: adapting AI models to customer data and domains, without code.
From research to the field
Our engines are optimized to run on a single compact appliance, so organizations can use state-of-the-art AI where the data is, even in disconnected environments.
Work with our research team
Have a domain, language or data type we should adapt to?
Research projects and grants
Funded research behind ZETTA AI Box
European Defence Fund (GA 101168045)
ATLAS
AI-based Natural Language Processing of Low-Quality and Multilingual Data in Defence Applications with User Adaptation
NLP systems resistant to noisy, multilingual defence data, covering speech and document recognition, translation, information extraction and user-driven adaptation.
European Defence Fund (EDF 2025)
MIDAS
Middleware for Intelligent Defence AI Dialogue Systems
A secure, explainable AI middleware that uses LLMs and AI agents for real-time decision support across strategic, tactical and operational levels.
EU Internal Security Fund – Police (GA 812631)
Do One Brave Thing
Do One Brave Thing (D.O.B.T.)
Countering radicalisation and online extremism among young people. ZETTA built BRAVE, a browser extension for verifying fake news and hate-speech content.
UEFISCDI
SOL2
Distributed Ledger Technologies in Internal Organizational Workflows and Support Applications for Cooperation among National Security Institutions
A secure multimedia data platform on blockchain infrastructure, with AI for entity recognition, translation and risk assessment.
UEFISCDI
SOL10 – RoNLP
Package of Tools for Processing and Linguistic Analysis of the Romanian Language
A toolkit for Romanian-language NLP, including sarcasm and irony detection and toponym geolocation.
ADR Nord-Vest
AI4INO – InoSafeAI
Multimodal Platform for Urban Public Safety
Multimodal analysis of video, images, voice and text to detect, classify and correlate events relevant to public safety, in a non-intrusive, explainable way that complies with the AI Act and GDPR.
Erasmus+ Alliances for Innovation (GA 101246679)
VITAL
Vocational Innovation and Training for Animal Livestock
An AI-powered decision support system pilot for the livestock sector, combining a conversational assistant with continuous data analytics to turn skills, training and pilot data into guidance for farmers, trainers and policymakers.
Research papers
Selected publications
IEEE ICCP 2026, Cluj-Napoca
Traceable Video-to-Text Generation for Multimodal OSINT Using Sparse Frame Sampling and Cross-Modal Evidence Alignment
George Bara, Emil Ștețco, Adrian Groza, Dorin Mirulescu
Generating video descriptions for OSINT that can be traced back to evidence in specific frames.
IEEE COMM 2026 – RISS Workshop, Bucharest
Transformer-Based Detection of Satire, Sarcasm, and Irony in Romanian Text
George Bara, Emil Ștețco, Daniel Bran
Transformer classifiers for figurative language in Romanian.
IEEE COMM 2026 – RISS Workshop, Bucharest
Semantic and Relational Toponym Geolocation from Romanian Texts: An LLM-Driven, on Premise Architecture with Ambiguity-Aware Disambiguation
Emil Ștețco, George Bara, Daniel Bran
On-premise, LLM-based geolocation of place names in Romanian text, including ambiguous references.
TRIZ Future Conference 2022, Springer IFIP AICT vol. 655, pp. 165–177
An Interactive Artificial Intelligence System for Inventive Problem-Solving
Stelian Brad, Emil Ștețco
An AI assistant for systematic innovation using the TRIZ method.
CogSIMA 2021
Building a Dynamic Corpus of Fake News Using Commercially Available Machine Translation and NLP Software
George Bara
A method for building a multilingual fake-news corpus with commercial MT and NLP tools.
Studia UBB Digitalia 65(1), pp. 37–52
Preliminary Research on Computer-Assisted Transcription of Medieval Scripts in the Latin Alphabet using AI Computer Vision techniques and Machine Learning
Adinel C. Dincă, Emil Ștețco
Computer vision for segmenting and transcribing medieval handwritten texts.
Springer, Information Quality in Information Fusion and Decision Making, pp. 181–206
Fake or Fact? Theoretical and Practical Aspects of Fake News
George Bara, Gerhard Backfried, Dorothea Thomas-Aniola
Theory of fake news and practical approaches to detecting it.
