01
IRIS
A vision–language model built around Ouro.
500MParams
4Training runs
2026Release target
hey I'm
David Seyser. I build models
that think in loops.
Researcher & MSc candidate at FHTW, working on a from-scratch vision–language model, regularization for post-training, and recurrent transformers that reason by iteration instead of by scale.
- Looking for
- Full-time ML research or engineeringIndustry research labs, applied science, or small focused teams.
- Start
- After my MSc wraps 2026FHTW, Vienna. Thesis on looped language models.
- Based in
- Austria · open to remote EU or relocation
- Stack
- PyTorch · JAX · CUDA · Python · C++
✓ usually reply in under 48h
research
Three things I'm pushing on.
02
LoopLMs
Recurrent transformers that can think for longer when a problem is harder. Same parameters, variable depth, a small model that decides to compute more. My bet on what comes after raw scale.
8×Effective depth
−2.1Loss vs. baseline
2026Paper target
03
ReSiReg
-
-
CoRLIn Progress
shipped
What I did.
2025
SOMP Stiffness Optimized Motion Planning
Workshop · peer-reviewed
Published
2026
ReSiReg regularization for stable post-training
CoRL
Under review
2026
IRIS
HuggingFace model
Published
SS '26
Course · Moderne Verfahren zur Sensorbasierten Roboteregelung
FHTW · MSc · 12 weeks
Teaching
WS '25
Course · Robot Based Manufacturing
FHTW · MSc · 14 weeks
Teaching
say hi
Let's actually
build tomorrow.
Hiring, collaborating, or just want to nerd out about loops, email's fastest. I read everything.
© 2026 David Seyser · Vienna, AT
Today for a better tomorrow.