I do self-driven AI research and test small language architectures. I am a fan of hybrid models, state-space dynamics, and gating mechanisms.
- Architectures: Hybrid Language Models (SSMs, Hopfield Networks, Latent Transformers)
- Time-Series: Autoregressive Simulation Modeling of Astronaut Stress data (including Temporal Fusion Transformers)
- Behavioral Dynamics: Time-Series Modeling of Psychological Behavior
- AI Safety: Value-Aligned AI, Transparency, and Democratic Governance in RLAIF/RLHF
- 🎸 GearChat - RAG-powered chatbot for musicians to chat with their instruments' manuals (synthesizers, drum-machines, etc.).
-
- Description: Native MLX reimplementation of the Mamba-3 architecture optimized for Apple Silicon (Metal). Replicates upstream block-selection logic, core Mamba-3 recurrence concepts, inference-cache semantics, and config structure, supporting mixed-precision SFT (LoRA) and hybrid Attention-Mamba models.
-
🌌 Persephone / VLM ([current reasearch])
- Description: Hybrid Vision-Language architecture merging a vision patch compressor, Mamba-3 linear scan, and Sparse MoE, featuring parallel agent contemplation over an internal Reasoning in Memory latent workspace.
-
- Description: A hybrid implementation of Shao et al. their Continuous Autoregressive Language Modeling (CALM) fused with State Space Models (SSM) and Hopfield Networks and gating + rectified flow.
-
🧠 Hamba
- Description: Novel hybrid Hopfield + Mamba inspired architecture combining SSMs and Hopfield networks with attention (BPE PPL of 16.2 on FineWiki), two hierarchical reasoning passes (Segmented-Reasoning), and rectified flow.
-
🏛️ Master's Thesis: RLAIF/RLHF for Public Value Alignment (REPO)
- Description: Enhancing transparency in LLMs through value alignment. Focused on aligning Mistral-7B with democratic transparency values (VU Amsterdam, 2025).
-
📊 Time-Varying VAR/GAM and Temporal Fusion Models
- Description: Extension of my BSc thesis extrapolated to astronauts on a 1-year mission. Predicting psychological behavior from simulated Astronaut stress data over a year with 3% shock event data (VU University Amsterdam, 2024).
- 📊 BSc Thesis: Estimating Emotions with Time-Varying-GAM and -VAR on High-Dimensional EMA data
- Description: Predicting psychological behavior from EMA data using Vector Autoregression (VAR) and Generalized Additive Models (GAM) (Maastricht University, 2022).
"Technical sophistication alone cannot resolve fundamental tensions in democratic governance"
— From my master's thesis on AI transparency