Context

For OpenBooks, I designed and built an AI reading tutor that models phonics sounds, listens to student responses, and gives pronunciation feedback. The question was whether that short exchange could produce useful guidance in a real teaching context.

Design the learning exchange

The tutor needed to model a sound, listen to a student attempt it, and respond at the phoneme level. I built the adaptive diagnostic and tutoring workflows around that interaction, with feedback that responds to the student’s attempt.

Evaluate with tutors

A model’s score needed a human reference. We evaluated the system against tutor-observed ground truth in live student lessons, connecting the technical feasibility question to OpenBooks’ teaching context.

I weighed per-task accuracy against inference cost when selecting speech models. The evaluation informed the client’s decision to continue investing in the approach.

What I learned

A convincing prototype was only part of the work. Comparing its behavior with tutor observations made the quality discussion concrete and gave the client evidence for the next decision.

Evaluation was part of designing the product: it shaped how we judged the interaction and which technical choices were worth pursuing.

My role

I designed and built the diagnostic and tutoring workflows and evaluated speech models for the task. OpenBooks was the client; tutors contributed the observed reference against which the system was assessed.

About this summary

This project was completed under NDA. Client interfaces, recordings, student data, and private evaluation materials are not shown here. This story covers my role and the public outline of the process.