Socratic: an adaptive AI learning companion

Students in underserved public schools often face a hard choice: work alone without timely help, or use an answer machine that completes the thinking for them. Socratic is built for a different kind of support. It guides a student toward an answer through one focused question at a time, while keeping the student in control of the reasoning.

The learning problem

A student who receives a final answer may finish an assignment but still carry the same misconception into the next lesson. Socratic helps with math, science, writing, and logic by diagnosing what the student currently understands, shrinking the next step when they are stuck, and asking them to verify their own reasoning.

How the agent adapts

Socratic maintains a structured student model on every tutor turn. It records:

progress earned from demonstrated understanding, rather than number of messages
confidence, frustration, and the depth of support used
misconceptions that are unresolved, resolved, or need another check
student-earned insights for a later recap

The agent uses that model to decide its next action: ask a diagnostic question, challenge an incorrect assumption, give a smaller hint, or ask the student to explain why a step works. A live Tutor's Notebook makes this adaptation visible, and the student can review their own learning history in My Progress.

Built for responsible learning

Socratic refuses requests for a direct answer or help during a live exam, then redirects the student to a safe next step or future practice. It treats prompt injection and answer extraction as learning-risk signals, not instructions. The goal is to help students build independence, not make copying easier.

The app is available in a mobile-friendly browser without an account or personal profile. Sessions can resume on the same device, and anonymous session reports show the thinking the student demonstrated rather than exposing a chat transcript.

Impact for underserved learners

A student does not need scheduled tutoring to receive patient, personalized practice. Socratic makes each interaction evidence-based: it adapts to a learner's misconception, preserves the student's agency, and creates a compact recap of what they actually figured out. This gives students a private, low-friction companion for building foundational skills and gives educators or families a shareable view of learning progress.

Technical approach

The deployed Next.js application uses Fireworks AI through JSON-schema structured output, so each model response contains a tutor reply plus validated student state. The UI uses that state to drive the hint ladder, Tutor's Notebook, recap, session report, and longitudinal on-device progress. The project has been manually stress-tested against answer begging, prompt injection, active-exam requests, wrong answers, storage corruption, and network failures.

Live app: https://socratic-tutor-ashy.vercel.app/

Source: https://github.com/JrKrishh/socratic-tutor

Demo: https://youtu.be/fhJD_rdA1dI