Akinola Olujobi

THE PROBLEM When a child struggles to read, the adults around them usually know something is wrong, but not why, what to try, or whether it worked. Most schools worldwide have no specialist to ask, and professional assessment can take months or may not be available at all. Teachers are left guessing and repeating the same practice, and a struggling child can be mistaken for a child who isn't trying. The same poor reading can come from a gap in prior knowledge, unfamiliar language, the format of the task, tiredness, or something that needs a professional. A wrong answer alone does not say which. THE SOLUTION MINDTRACE is an AI agent that helps teachers investigate why a learner struggles to read, choose a support, and verify whether it worked. It gives educational observations, not diagnoses or medical advice, so a teacher can act today with evidence instead of waiting. HOW IT WORKS 1. Observe. The teacher records a concern in their own words. No condition is asked for or named. 2. Evidence. A child reads a short passage aloud with an adult present, and the teacher enters what was heard. Code (not a language model) aligns every word into the Trace: read as written, read differently, or skipped. Every word has its own evidence ID, and every count can be checked. 3. Investigate. An agent built on Muse Spark reads the evidence through tools, weighs competing explanations (including language background, format and recording quality), and drafts a report. A separate verifier and deterministic filters check each claim against the evidence and block diagnoses, probabilities and treatment advice. 4. Support and verify. The teacher picks a concrete strategy, such as sound-by-sound pointing, and MINDTRACE assigns a matched follow-up passage. The comparison is worded so it never claims the support caused the change. WHO BENEFITS Teachers and tutors in under-resourced schools, anywhere, who have no specialist to ask. It is designed for different languages and regions, because reading difficulty looks different across writing systems, and it says when an assessment has not been validated for a learner's language. PRINCIPLES Evidence before conclusions, educational observations not diagnoses, the teacher stays in control, and fail closed: if the model is unavailable, the evidence is kept and nothing is invented. Every learner in the demo is synthetic. META MODELS Muse Spark 1.3 powers the investigation agent and verifier. [CHOOSE ONE: (A) The investigation runs live in the demo. / (B) The model key was unavailable at recording time, so the investigation is shown failing closed. The provider layer, agent loop, verifier and tests are implemented, and real Muse Spark chat calls were verified earlier this week.] Speech-to-text for the read-aloud is planned but not part of this submission, so the teacher types the transcript. Built solo with Muse Code during the hackathon (Oct 27 to Nov 3, 2026). TECH Next.js, TypeScript, Tailwind, an OpenAI-compatible provider layer with a model allowlist, tested evidence, comparison and safety code, and a Supabase schema with row-level security (the demo currently uses a local store).