Marius-Constantin
Dinu.

A neuro-symbolic AI company selling verifiable agent workflows into regulated industries — legal, finance and the public sector — where an answer that cannot be checked is not an answer.

Founded it, set the technical thesis, and architected the system that made multi-step generation reliable enough to sell. Zero to over EUR 1M in revenue, profitable in the first year, and named by Sifted / Financial Times among Europe's top 11 AI startups to watch.

A production pipeline for game and film studios: concept art in, production-ready 3D assets out, with the generation steps governed tightly enough that a studio can put them in a real pipeline.

Directed the applied research agenda across deep learning and reinforcement learning, advised leadership on model selection and build-versus-buy, and built production-facing prototypes from problem framing through evaluation to hand-off.

One of Austria's principal research universities, and the home of the LIT AI Lab under Sepp Hochreiter — the group where LSTM originated.

Doctorate in Artificial Intelligence, passed with distinction, plus five years as a research scientist in the lab. Solved hyper-parameter selection in unsupervised domain adaptation, where no target labels exist to validate against: ICLR 2023 Oral, top 5%.

A European frontier lab building xLSTM, a recurrent architecture positioned as an alternative to the Transformer for long-context and edge workloads.

Co-author on "Large Language Models Can Self-Improve at Web Agent Tasks", where NXAI is a shared title-page affiliation with Markus Hofmarcher and Sepp Hochreiter.

The Johann Radon Institute for Computational and Applied Mathematics, part of the Austrian Academy of Sciences — inverse problems, numerical analysis and mathematical foundations.

Co-author on the SymbolicAI paper (CoLLAs 2024) with Werner Zellinger, who carries the RICAM affiliation and co-supervised the doctoral work on regularisation and the balancing principle.

An Ivy League research university whose NLP group, under Chris Callison-Burch, is among the longest-running in the field.

Co-author with Ajay Patel and Chris Callison-Burch on "Large Language Models Can Self-Improve at Web Agent Tasks" — self-improvement for agents acting in a live browser.

A software-intelligence platform for observability at enterprise scale: tracing, monitoring and automated root-cause analysis across large production estates.

Senior machine learning researcher at the LIT Open Innovation Center, building deep learning and NLP prototypes and working across research and product to move observability capabilities from paper to pipeline.

A digital asset management product used by large product- and brand-centric organisations to run the content supply chain behind their marketing.

Developed and deployed deep-learning image classification for automated tagging inside the product — the feature that turns an unlabelled asset library into a searchable one.

The corporate research arm of Siemens in Princeton, New Jersey, where applied research is taken toward product across the group's industrial businesses.

Built a cross-platform handwritten-character recognition app combining SVMs and neural networks, with a C++ inference core bridged to Xamarin and the Android NDK through native interop.

An embedded-computing and IoT manufacturer supplying industrial hardware and the software that runs on it across Europe.

Led an international team delivering embedded and server software for self-service coin-counting devices, and designed the SOAP integration layer spanning Java and C#/.NET. Five years, ending as software architect and product manager.