Selected work

Projects at the edge of models and people.

Two close looks at how I frame problems, build systems, and learn from the limits of data.

Research prototype · 2026

Turning wrong answers into useful teaching signals

A learning-support prototype that classifies incorrect mathematics responses using a 14-category misconception taxonomy—then delivers hints that guide without giving the solution away.

The question

A wrong answer tells an educator that something went off track, but rarely explains where. The project asks whether a structured misconception model can make that error legible to students, teachers, and families.

The approach

I built the React and TypeScript prototype across Algebra, Geometry, Calculus, and Linear Algebra. Student-facing feedback emphasizes the next reasoning step, while instructor and parent views translate response patterns into heatmaps, intervention priorities, and plain-language explanations.

The work reinforced a principle I return to often: a prediction becomes valuable only when it helps someone decide what to do next.

EECS 486 capstone · 2026

A more honest view of social media risk

A machine-learning and information-retrieval project exploring self-reported productivity loss, with an interactive experience that compares a user with similar profiles.

The question

Survey data about attention and productivity is noisy, subjective, and often internally inconsistent. A polished score can hide that uncertainty rather than clarify it.

The approach

I evaluated an XGBoost model on a public dataset, then studied where data quality constrained performance. The companion web application uses vector embeddings to find similar profiles and presents comparisons instead of treating one prediction as a universal verdict.

The strongest outcome was not a single metric. It was a product direction that surfaces context and uncertainty so users can interpret a model’s result responsibly.