Emanuele Di Rosa PhD · CTO · INVENTOR
WORK & RESULTS

Strategized, designed, built, benchmarked, patented, shipped.

Five pieces of work that show the whole arc — research, invention, product, and leadership — with the numbers behind them. Benchmarks are stated as measured; internal metrics are labelled as such.

CASE A · AGENTIC AI · PRODUCTION · DATAmundi

AIDA Agents: a multi-agent LLM platform that beats state of the art

Challenge. Enterprise localization faced a trade-off: NMT engines need expensive fine-tuning and heavy post-editing; LLMs promise quality but lack workflow integration.

What I did. Led the research team behind AIDA Agents and its implementation: an orchestration of specialized LLM agents — translator → arbiter/rater → post-editor → re-rater — with optional RAG injecting translation memories, terminology and style at every stage, an MQM-inspired explainable quality report, three quality/cost tiers, and native XLIFF/CAT-tool integration. No fine-tuning.

Result. Deployed in production for enterprise clients; won DATAmundi a Globee Innovation Award in the Best Multi-Agent System Orchestration Platform category; a 2026 AI Excellence Awards finalist; published at ACL 2026 (Industry Track) and EAMT 2026.

Verify: Globee award →AI Excellence finalist →ACL 2026 paper →EAMT 2026 paper →Slator coverage →
10/11WMT24++ language pairs where it beat every system, with no TM or terminology
70–98%publication-ready segments on an industrial benchmark, no human post-editing
99.4%terminology accuracy — beat all 20 systems on the WMT25 terminology task
50–70%potential reduction in post-editing effort (EAMT 2026 estimate)
CASE B · AI SAFETY · RESEARCH → SYSTEM

Formal-logic red-teaming and the AI safety-auditing system

Challenge. Most red-teaming is ad hoc and non-reproducible, and single-metric evaluation over- or understates risk — a fatal flaw for auditing.

What I did. Designed a deterministic method that encodes sensitive queries into first-order logic to reveal where safety alignment silently breaks; a dual-metric framework separating format engagement from genuine commitment; LLM-as-judge validated against human annotation; provenance-hashed, exactly reproducible runs. The method is now the core of an AI safety-auditing system I lead at DATAmundi, with a PhD student from the Georgia Institute of Technology.

Result. Accepted at the PANDORA workshop, EMNLP 2026; an extended multilingual version under review.

Verify: PANDORA workshop →OpenReview profile →
14frontier LLMs from 4 providers, paired experiments
37.2%safety-bypass rate, robust across reconciliation methods
672paired observations across three model generations
SHA-256provenance-hashed prompts, exactly reproducible runs
CASE C · INVENTION · ISSUED PATENT

Provably optimal recommendation under hard constraints

Problem. Recommendation and search rank by heuristics; they cannot guarantee that hard requirements are respected or that the top result is actually optimal — and they cannot explain why.

Invention. A method that represents mandatory constraints and partially-ordered qualitative preferences as a DAG, assigns weights with a separation guarantee (no combination of lower preferences can outweigh a single higher one), and returns candidates with mathematical guarantees of optimality and equivalence — incrementally, without restarting as new sources arrive. It applies to search ranking, RAG retrieval, shopping co-pilots, and automated negotiation.

Result. US Patent 12,717,799 B2, issued Aug 25, 2026, and Italian Patent 102024000006643, granted Jul 1, 2026 — solo inventor. The engine behind PerfectFinder.ai.

Verify: Google Patents →Certificates →
Italian patent certificate n. 102024000006643
STATUSIssued — US 12,717,799 B2 (Aug 2026) · IT 102024000006643 (Jul 2026) · solo inventor
WHY IT MATTERSConstraint-aware, explainable, provably-optimal recommendation is a large, under-served space; the same core improves RAG retrieval and agentic decision-making.
ALSOTrueFile.ai — decentralized verification of digital-content authenticity and integrity, with similar-content detection. Patent pending in multiple countries.
CASE D · SAAS · App2Check

From research to two SaaS platforms licensed to global brands

Context. A product venture within the Finsa software group. I acted like a co-founder, with the freedom the owners gave me, and led product and all R&D.

What I built. The App2Check Voice-of-the-Customer analytics platform — ML/NLP sentiment and semantic analysis across app reviews, social, chats and email — and EVA, a no-code cognitive-chatbot platform that turns FAQs and documents into a domain chatbot through patented knowledge-base extraction.

Result. Licensed to leading brands including Stellantis, Adidas, TIM, UniCredit and Poste Italiane; selected by Precisely (formerly Pitney Bowes software) as a worldwide partner, with EVA integrated into EngageOne Converse; winner of all three ATE_ABSITA tasks at EVALITA 2020.

Verify: Pitney Bowes partnership →EVA talk, Chatbot Summit 2019 →EVALITA 2020 results →Patent (Justia) →
Emanuele Di Rosa presenting App2Check’s sentiment-analysis results on 10,000 app reviews, compared with other tools
“
Emanuele applies his knowledge of software engineering and machine learning to address meaningful business challenges. He generates creative solutions to difficult problems, and always accounts for client needs and the customer experience. A great professional to have as a business partner.
Christopher Cummings — Director of Product Management, Precisely · Verify on LinkedIn →
CASE E · CTO TRANSFORMATION · DATAmundi

Turning a technology function into a product-and-R&D engine

Challenge. Turn a services company’s technology function into one that ships products and publishes research, without letting delivery crowd out the roadmap.

What I did. Set AI product strategy and an R&D roadmap that connects research to research and technology grants, funding and partnership shaping with US and EU universities; led the modernization of the two SaaS platforms — AIDA Hub, for managing data-services projects, including through AI agents, and DATAtalent, for contributors and subject-matter experts; led security, compliance and the modernization of finance and operations flows; ran AI discovery with clients that are primarily big-tech companies, across several of their research and product departments, including frontier labs.

Result. A research-to-product engine: ACL 2026 and EAMT 2026 papers, a Globee Innovation Award and an AI Excellence finalist place for AIDA Agents, and partnerships shaped with US and EU universities.

Verify: CTO appointment →ACL 2026 event page →
2peer-reviewed papers from the R&D line: ACL 2026 and EAMT 2026
2SaaS platforms modernized: AIDA Hub and DATAtalent
US · EUuniversities partnered with; research and technology grants and funding shaped
1 + 1Globee Innovation Award won; AI Excellence Awards finalist