Back to main portfolio

AI in Education

This portfolio explores how AI can support learning without hiding uncertainty, weakening learner agency, or replacing professional judgment. The projects focus on observable learning processes, transparent analytics, responsible decision support, and research designs that can be challenged and extended.

Generative AI Learning Analytics Human and AI Collaboration Responsible AI Explainable AI Multimodal Analytics Privacy Preserving ML Process Mining Classroom Discourse Collaborative Learning

Research Prototypes

These repositories are working research prototypes rather than claims of completed empirical studies. Their demo outputs use synthetic data unless the repository states otherwise. Each project keeps assumptions, metrics, code, tests, and limitations visible so the next step can be real validation rather than a polished black box.

Generative AILearner agency

GenAI Learning Observatory

  • Analyzes learner interaction with generative AI through verification, revision, reflection, and adoption behavior.
  • Keeps event data, feature extraction, aggregation, and reporting separate so each signal can be traced to the underlying interaction.
AssessmentHuman review

Teacher and AI Assessment Studio

  • Studies rubric scoring as decision support rather than automatic judgment.
  • Separates AI scores, teacher scores, disagreement, uncertainty, routing, and final resolution for auditable review.
PrivacyLearning analytics

Privacy Preserving Learning Analytics

  • Compares centralized, federated, and noisy training on the same synthetic learner prediction task.
  • Makes privacy and utility tradeoffs visible instead of treating privacy as a footnote to model performance.
Multimodal AnalyticsSelf regulated learning

Multimodal Self Regulation Lab

  • Explores interaction, attention, self report, missing modalities, ablation, and reliability aware feature fusion.
  • Preserves modality level visibility so researchers can see when a signal helps, hurts, or disappears.
Process MiningInstructional design

Learning Design Process Mining

  • Treats learning design work as an event log rather than only a finished artifact.
  • Measures trace variants, transitions, entropy, rework, bottlenecks, and human and AI handoffs.
Hybrid IntelligenceDecision support

Hybrid Intelligence Lab

  • Tests when a human, an AI system, or a shared process should make or defer a decision.
  • Separates confidence, disagreement, routing, override, and deferral so complementarity can be studied directly.
Explainable AIFaithfulness

Explanation Faithfulness for AIED

  • Tests whether an educational AI explanation follows the model behavior it claims to explain.
  • Implements comprehensiveness, sufficiency, stability, and a combined faithfulness view while keeping plausible explanations separate from faithful ones.
Collaborative LearningDiscourse analytics

Collaborative Reasoning Analytics

  • Studies semantic uptake, participation balance, evidence use, questioning, and reasoning moves in group dialogue.
  • Keeps these signals separate so a collaboration summary can be inspected rather than reduced to one opaque score.
Generative AICognitive offloading

Cognitive Offloading Analytics

  • Turns possible cognitive offloading into transparent behavioral hypotheses using revision, verification, recall, confidence change, and similarity signals.
  • Treats the resulting score as an analytic proxy rather than a diagnosis of what a learner is thinking.
Classroom AnalyticsTeacher reflection

Classroom Discourse Intelligence

  • Extracts interpretable features from classroom dialogue, including pedagogical moves, questioning depth, feedback, learner elaboration, and uptake.
  • Frames the output as support for reflection and research, not as an automatic judgment of teaching quality.

Research Principles

Traceable measures

Metrics should be connected to observable data and assumptions that another researcher can inspect.

Human judgment remains visible

AI is treated as a source of evidence or decision support, not as an unquestioned replacement for teachers, learners, or researchers.

Synthetic data are labeled

Demonstration outputs are clearly separated from empirical findings. Synthetic results are useful for testing a method, not for claiming an educational effect.

Reproducibility before complexity

Small auditable baselines make it easier to test assumptions before moving to larger models, real institutional data, or higher stakes deployment.

Contact

For AI in Education research collaboration, doctoral discussions, or questions about these prototypes, you can reach me through the channels below.