Guus Bouwens

ML Systems · Interpretability · Research

Data Scientist &
AI Engineer

I design and ship production ML systems — LLM and RAG pipelines, recommendation engines, and evaluation harnesses — and research mechanistic interpretability: how internal representations emerge across training regimes, modalities, and levels of grounding.

Currently building Re-Search and a Nutrition Scanner. MSc Data Science, top of class.

3Peer-reviewed papers
1M+Chunks in production RAG
4.0MSc GPA
View my work Let’s connect Resume ↗

01 Research

Research

Layer-Wise Category Structure in LLMs TMLR

Found broad early-to-late representational organization across architectures, and showed fine-grained layer rankings are statistically unstable.

Causal State Variables in V-JEPA 2 Latents ICML

Causal interventions on video models revealed early-layer motion encoding and new portability metrics for causally privileged subspaces.

Document Structure Subspaces in OCR Models ICML

Specialized OCR models show far higher representational modularity and compression than general-purpose vision-language baselines.

02 Projects

Selected Projects

Local RAG

Production-grade retrieval over 800+ PDFs (1M+ chunks): hybrid search, cross-encoder reranking, streaming, and query rewriting at 6–8s/query.

Python · ChromaDB · BM25 · FastAPI Local RAG

Food → Nutrients App

AI pipeline turns food photos into macro/micro nutrient breakdowns via automated USDA matching. Live demo.

Computer Vision · YOLO · USDA API Food to Nutrients App

Crypto Sentiment Dashboard

Live multi-source sentiment over the top 250 coins — news RSS, Reddit, Bluesky, Mastodon, and YouTube titles via cron — blended with Fear & Greed and market context.

PHP · Cron · RSS · Bluesky · Reddit Crypto Sentiment Dashboard

Mechanistic Interpretability

Probed DeepSeek-R1 layer activations across tasks to reveal specialization and emerging cognitive patterns.

PyTorch · TransformerLens · Probing Mechanistic Interpretability

Highway Driving (RL)

Full top-down driving stack with reward-shaped PPO; survives 200-step evaluations, high-speed and collision-free.

Python · Stable-Baselines3 · PPO Highway Driving RL

Pulse Rate Algorithm

PPG + accelerometer signal pipeline with RF regression: 8.8 BPM MAE at 90% availability, validated on clinical CAST data.

Python · Signal Processing · RF Pulse Rate Algorithm

University (grade)

Professional

Independent

03 Certificates

Certificates

AI

Data Science

Analysis

Finance

Health

Other

04 Writing

Blogs

How I Built a Research Assistant

The Proxy Trap in AI Training

From Imitation to World Models

Claude's Architecture

Diffusion LLM Mercury 2

Prompt Repetition: A Surprisingly Simple Way To Boost LLM Accuracy

Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers

Alphabet’s Ascent to the Top

What Pulled Me Into the AI

LLMs explained: mathematically and in plain english

Paper: LLMs are Invertible

Thought: Google Should Win

Thought: Are We Summoning Ghosts or Building Animals? (RL vs LLMs)

Machine Learning for Healthcare - MIT 6.S897

Building a High-Quality RAG Application

DL 2021-2025 Research Papers

DL 2018-2021 Research Papers

Ilya Sutskever Research Papers 3

Ilya Sutskever Research Papers 2

Ilya Sutskever Research Papers 1

Deep Learning Algorithms 4: Generative & Advanced Architectures

Deep Learning Algorithms 3: Sequence Models

Deep Learning Algorithms 2: Computer Vision Architectures

Deep Learning Algorithms 1: Fundamentals

Machine Learning Algorithms 4: Ensemble Methods

Machine Learning Algorithms 3: Reinforcement Learning

Machine Learning Algorithms 2: Unsupervised Learning

Machine Learning Algorithms 1: Supervised Learning

Deep Learning for Computer Vision - Stanford CS231N

Language Modeling from Scratch - Stanford CS336

Foundation Models and Generative AI - MIT 6.S087

Reinforcement Learning - Stanford CS234

Machine Learning - Stanford CS229

Algorithms - MIT 6.006

Mathematics for Computer Science - MIT 6.1200J

Statistics for Applications - MIT 18.650

Fine-Tuning LLMs for Investment-Grade Company Classification

05 Contact

Let’s build something.

Reach out on LinkedIn or send a message below.