CV
Education, experience, publications, and skills.
Contact Information
| Name | Stein Pleiter |
| Professional Title | Chief Technology Officer & Co-founder, eBrain.ai |
Professional Summary
CTO & co-founder at eBrain.ai, AI RLHF Training Director at 10x.Team, and BSc Informatica candidate at the University of Amsterdam researching mechanistic interpretability of Mixture-of-Depths routers.
Experience
-
2026 - present Amsterdam, The Netherlands
AI RLHF Training Director
10x.Team
-
2025 - present Amsterdam, The Netherlands
Co-founder & Chief Technology Officer
eBrain.ai
Building an AI-powered operations platform for European SMEs.
-
2024 - 2025 Amsterdam, The Netherlands
Co-Founder
Hulpmetdewet.ai
AI legal assistant for Dutch lawyers and paralegals.
-
2024 - 2024 Amsterdam, The Netherlands
Full-stack AI Developer
AI Assist
-
2023 - 2025 The Netherlands
Entrepreneur / Developer
Tag 31
Education
-
2023 - 2026 Amsterdam, The Netherlands
BSc
University of Amsterdam
Informatics
- Bachelor’s thesis: “The Router’s Mind: Diagnosing Mixture-of-Depths Routers with Sparse Autoencoders” — graded 9/10.
-
2025 - 2025 Toronto, Canada
Awards
-
2026 Bachelor's thesis graded 9/10
University of Amsterdam
“The Router’s Mind: Diagnosing Mixture-of-Depths Routers with Sparse Autoencoders”, supervised by Dr. Ana Lucic and Ege Erdogan, examined by Dr. Dolly Sapra.
-
2026 Accepted virtual poster, ICML 2026 Workshop on Mechanistic Interpretability: "What Do Mixture-of-Depth Routers Learn? Routing Patterns in Gemma 2"
ICML 2026 Workshop on Mechanistic Interpretability
Companion paper to the bachelor’s thesis, on part-of-speech-dependent routing behaviour in Mixture-of-Depths models.
Publications
-
2026 The Router's Mind: Diagnosing Mixture-of-Depths Routers with Sparse Autoencoders
University of Amsterdam (Bachelor's thesis)
Uses sparse autoencoders and causal interventions to characterize what Mixture-of-Depths routers read from the residual stream in Gemma 2 2B.
-
2026 What Do Mixture-of-Depth Routers Learn? Routing Patterns in Gemma 2
ICML 2026 Workshop on Mechanistic Interpretability (virtual poster)
MoD routers trained on Gemma 2 2B prefer full-attention to sliding-window layers and route by part-of-speech in a layer-dependent way.