Wels · Austria — neuro-symbolic AI, verified systems
Marius-Constantin
Dinu.
I build AI systems that are held to a specification rather than trusted on output. A language model, used well, is a semantic parser: it turns intent into typed structure that a solver can actually check.
PhD in Artificial Intelligence under Sepp Hochreiter. Founder of ExtensityAI. Author of SymbolicAI. Currently building governed agent infrastructure at Alpha Omega Labs.
- EUR 1M+ Source: ExtensityAI, founded 2023 with no prior revenue: scaled past EUR 1M and profitable within its first year of trading, on a EUR 767k pre-seed. Founder record.
Trajectory
Research and shipping, at the same time — not in sequence.
Most CVs flatten a career into a list, which hides the part that matters. The doctorate, the industry research post, the company and the advisory role were not phases. They ran together.
- Kontron
- Siemens Corporate Research
- CELUM
- Imagine Kara LLC
- Johannes Kepler University — LIT AI Lab
- Dynatrace Research
- ExtensityAI FlexCo
- Atlas
- Alpha Omega Labs
Systems
Things I built and still run.
Each of these is in production. The numbers come from the repositories, measured, not estimated.
-
Cortex
A governed control plane for AI agents
A private control tower that lets a company run AI coding agents from chat apps like Slack or Telegram — with rules about who may do what, a full audit trail, and one-click deployment of whatever the agents build.
- 182,600 lines
- 53
- 643
-
omegaXiv
Autonomous research platform
An online platform where AI agents actually run a scientific investigation end to end — planning it, executing code in isolated sandboxes, checking their own work against strict quality contracts, and publishing a finished, peer-reviewable paper.
- 613,600 lines
- 8,100 Python + 3,900 TS
- 9
-
SymbolicAI
Open-source neuro-symbolic framework
An open-source framework that combines language models with mathematical solvers, so a complex task breaks into small steps that can each be checked for correctness instead of trusted.
- 1,745
- 91
- 49
Research
380 citations, and one recurring argument.
Published at ICLR, NeurIPS, ICML and CoLLAs. The through-line is verification: how to choose a hyper-parameter when you have no labels to validate against, how to assign credit from two demonstrations, how to make a generative model's output checkable rather than plausible.
8
8
ICLR 2023 Oral (top 5%)
What others say
-
The work is brave (very much challenging a dominant paradigm), and novel. Without hesitation, I give it my highest recommendation.
Source: Published on the ExtensityAI website. Marcus was external examiner of the doctoral thesis.
Gary Marcus
-
What distinguishes Dr. Dinu is his remarkable ability to translate theoretical advances into practical impact while maintaining strong academic collaborations.
Source: Excerpt from a signed recommendation letter, 9 December 2024. Hochreiter supervised the doctorate.LetterSepp Hochreiter
-
He consistently demonstrated an ability to look beyond current paradigms, envisioning and implementing solutions that anticipate future developments in the field.
Source: Excerpt from a signed recommendation letter, 9 December 2024. Zellinger was doctoral co-supervisor.LetterWerner Zellinger
-
Interesting reading for those that are looking at the evolution of Symbolic AI methods.
Source: Published on the ExtensityAI website, from a LinkedIn post about the SymbolicAI paper.
Pietro Leo