StudyStride ReasonTrace AI

Inspiration

Most learning systems celebrate a correct answer. But a student can reach the right option through an invalid rule—and that misconception survives until the exam. StudyStride ReasonTrace AI asks a second question: how did you get there?

What it does

ReasonTrace is a private, local-first SAT reasoning assistant that detects the first divergence between a learner's answer and their reasoning path.

In the demo, a learner solves 3x + 5 = 20, selects the correct answer, and then reveals the step they used. If they choose “divide only the variable term,” ReasonTrace identifies the hidden misconception: right answer, wrong reason.

The AI operates as four coordinated roles:

• Observer captures the selected answer, reasoning path, timing, and confidence.
• Skeptic checks whether the stated reasoning is compatible with the algebra.
• Diagnostician maps the first divergence to a misconception graph and updates mastery.
• Coach generates one precise next step and an adaptive seven-day practice plan.

Every diagnosis is inspectable. The AI Summary shows the evidence it observed, its inference, confidence, and recommended next move—so learners can understand why the system intervened.

How we built it

The prototype uses a Python standard-library server with a responsive HTML/CSS/JavaScript interface. It runs locally without an account and does not collect personal information.

The intelligence layer is purpose-built for traceable learning decisions:

• Bayesian Knowledge Tracing estimates skill mastery after each observation.
• A misconception graph connects reasoning patterns to likely conceptual gaps.
• A diagnostic policy finds the first invalid step instead of merely grading the final answer.
• A utility-based task selector chooses the next practice item from mastery, uncertainty, and expected learning value.

This is symbolic and probabilistic AI rather than a hidden chat response: each recommendation is derived from visible evidence and can be inspected in the interface. OpenAI Codex assisted with product design, implementation, testing, and demo production.

Challenges we ran into

The hardest problem was separating answer correctness from reasoning correctness without pretending that the system understood evidence it never collected. We designed explicit reasoning-path interactions, confidence states, and a deterministic evidence trail. We also had to make a multi-agent architecture understandable in under three minutes while keeping the experience accessible and fast on low-end devices.

Accomplishments that we're proud of

• Detects a correct answer produced by an invalid rule.
• Surfaces the first reasoning divergence, not just a red score.
• Provides an evidence-backed AI Summary with calibrated confidence.
• Coordinates four distinct AI roles in one coherent learning loop.
• Generates a mastery-aware seven-day plan.
• Runs locally, requires no login, and keeps learner data on the device.
• Includes seven automated tests for the core diagnostic behavior.

What we learned

Correctness-only feedback can accidentally reinforce misconceptions. Asking for one lightweight reasoning choice creates much richer pedagogical evidence. We also learned that educational AI becomes more trustworthy when its evidence, inference, confidence, and next action are separated and shown explicitly.

What's next for StudyStride ReasonTrace AI

Next we would add free-form reasoning input, privacy-preserving teacher dashboards, more SAT domains, and lightweight on-device language understanding. The larger vision is an AI learning layer that rewards sound thinking—not lucky answers.

Built With

• accessibility
• bayesian
• knowledge
• offline
• python
• symbolic