Zion Krullaars — Research

Evidence-based interaction between humans and Artificial Intelligence: technical solutions co-designed with users, scientific human-in-the-loop experiments, and managing sustainable technical solutions.

BASED IN ZÜRICH

Portrait of Zion Krullaars — Photo · Zion Krullaars

Hi, I am Zion. My research treats interactive artificial intelligence as something to define, build, and test under clear assumptions—not just to prototype. I combine human-centered design (including human-computer interaction, human-in-the-loop, and co-design) with artificial intelligence so that all claims are backed by clear methods and evidence.

I hold a Master of Science in Interaction Design from Delft University of Technology (GPA: 8.1/10) and a Master of Science in artificial intelligence from Vrije Universiteit Amsterdam (GPA: 8.8/10). My graduation projects focused on learning-based artificial intelligence with teacher supervision. I used reinforcement learning to develop tutor dialogue strategies, which included tracking a student's mental state over long conversations.

I recently moved to Zürich. Coming from the flat Netherlands, I still stare at the mountains as if they just appeared for the first time. When I am not working, I am likely out on my bike, nerding out over coffee beans, or starting a project that is way to big for the time planned (and finish it).

This website is a brief record of my projects and research. My interests lie in artificial intelligence research, human-computer interaction, and reinforcement learning. I am especially interested in projects where human-AI interaction is a core part of the research problem itself, rather than just the interface. I am currently based in Zürich, Switzerland. For contact, please see my contact section.

Highlighted work · 03 Highlights

MSc thesis · Design for Interaction · TU Delft · 2026

Co-creating a human-centered AI learning system for the future of education

Co-Design · theme: high school education · product: Cubo · grade: 9/10

Problem: Generative AI in schools defaults to opaque student–model dyads: shortcuts to answers, weak teacher oversight, and hard-to-verify learning.
Method: Participatory co-design with a mixed educator–student think tank; three iterative cycles (exploration → prototyping → evaluation). Conceptual frame: flight-simulator roles (pilot / co-pilot / control tower). Artefact: web platform (student cockpit, teacher tower, curriculum upload, physical starter kit).
Evidence: Comparative field study, 10 days, n = 16 students vs generic AI tools: reported gains in confidence; misinformation-detection task 100% success vs baseline; teacher-rated monitoring and mentoring affordances (see image below and thesis for further instruments and statistics).
Limits: Single deployment context; self-report and task-based measures; generalisation to other subjects and scales requires further study.

Human centred AI classroom study interface overview

Teacher workflow and oversight interface for the human centred AI classroom study

Student workflow and progress interface for the human centred AI classroom study

Repository link → thesis + supplements

MSc thesis · Artificial Intelligence · VU Amsterdam · 2026

Think before you talk: pedagogical strategies for LLM tutors via reinforcement learning

AI Research · Theme: Reinforcement Learning · Post-training an LLM tutor · grade: 9/10

Problem: Instructionally sound tutoring (scaffolding, answer withholding) conflicts with LLM training objectives that favour immediate helpfulness; a prompted frozen model alone can obey surface rules yet produce dialogues where no learner reaches an independent solution.
Method: Modular tutor: a Message Interpreter maps the dialogue to features for a trainable StrategyDirector policy, which picks discrete pedagogical actions; a frozen Tutor LLM (Gemini 2.5 Flash) realises those actions in natural language. A simulated StudentSimulator (same model family) drives multi-turn practice. A Feedback Engine scores each tutor turn against explicit Teacher Rules and returns rewards to the policy, closing the RL loop.; Conceptually, teacher-aligned strategy and learner state inform the director, which issues a strategy prompt to the static generator. Optimisation over full trajectories uses PPO and GRPO vs a strong prompt baseline, in a single algebra tutoring MDP.
Evidence: Baseline lowest per-turn answer leakage (3.3%) but zero independent student solutions. RL policies yield higher independent solution rates (PPO 36.7%; GRPO 53.3%) with engaged student states. PPO maximises reward alignment (state–strategy 0.813; Socratic-dominated); GRPO shows lower alignment (0.567) but higher engagement (p < 0.001) and more diverse strategies.
Limits: Simulated learner; one task family; controlled conditions—external validity to real classrooms and broader curricula is open work.

Architecture overview for strategy-based reinforcement learning tutor

State and feedback dynamics in strategy-based reinforcement learning

Evaluation results for strategy-based reinforcement learning tutor

Thesis publication pending

Publication · Mensch und Computer 2023

Exploring Initial Interactions: High School Students and Generative AI Chatbots for Relationship Development

Interaction Study · Theme: High School Education · MuC 2023

Aim: Identify how high-school students form early relationships with generative AI chatbots, and derive practical recommendations for responsible classroom adoption.
Method: Two-day exploratory study using mixed qualitative signals (group activities, chatbot interactions, interviews, assignments, scale responses, and word-cloud elicitation), followed by synthesis into need-based student personas.
Findings: Students differ strongly in trust and dependency profiles; effective designs should personalise tutor behaviour, support skeptical learners, and frame chatbots as supportive learning partners rather than answer engines.
Citation: Krullaars, Januardani, Zhou, Jonkers (2023). Mensch und Computer 2023 — Workshopband, MCI Student Research Competition. DOI: 10.18420/muc2023-mci-src-415.

Visual summary figure from the MuC 2023 paper

Supplementary figure from the MuC 2023 paper

Publication link → GI digital library Local file → MuC paper (PDF)

Theme

Choose a skill pill to read details.

Course Portfolio

Assessment for Impact Bio-Inspired Design Context and Conceptualization Design Theory and Methodology Research Methodology Interactive Technology Design Product Understanding, Use and Experience Project Exploring Interactions Reflection on Designing Visual Communication Design AI and Society Data Mining Techniques Evolutionary Computing Experimental Design and Data Analysis for AI (Statistics) Human-Interactive Agent Learning Knowledge Representation Multi-Agent Systems Natural Language Processing Programming in Python for Text Analysis Skills for AI Socially Intelligent Robotics Development, Sustainability and Culture Fundamentals of Entrepreneurship Intercultural Internship Preparations for Intercultural Internship Research and Design Business Culture and Technology Concept Design Design Driven Innovation Design for Sustainability Designing Connected Experiences Embodiment and Detail Design Engineering for Design Ergonomics and Consumer Behavior Form and Experience Integrated Technology Interaction and Electronics Introduction to Industrial Design Engineering Manufacturing and Design Product Dynamics Product Statics Research and Design

Publications & artefacts

TU Delft MSc thesis + poster (institutional repository)
VU Amsterdam MSc thesis (publication pending)
MuC 2023 paper: Exploring Initial Interactions (GI Digital Library)

Contact

Write to zionkrullaars[at]gmail[dot]com or use LinkedIn. I am based in Zürich, Switzerland, with EU residence. Research interests: AI research, HCI, reinforcement learning, and embodied technology with human interaction.

CV LinkedIn GitHub