AccessLearn

"We don't just tell you that you're wrong. We help you understand why." Accessible AI companion featuring grounded explanations, misconception diagnosis, and adaptive retry loops.

Comment

Inspiration

Traditional educational tools and chatbots often operate as binary judges: they either tell you that you're "Wrong" or dump an unhelpful textbook answer. For neurodivergent learners, non-native speakers, or students grappling with dense STEM texts, this creates cognitive fatigue, anxiety, and learned helplessness.

We asked ourselves: What if an AI tutor acted like a master teacher? One that doesn't just evaluate whether you reached the right conclusion, but actually diagnoses why you arrived at a misconception, respects your intuition, and actively adapts its pedagogical approach until genuine conceptual clarity is achieved. This sparked AccessLearn — built around our core motto: "We don't just tell you that you're wrong. We help you understand why."

What it does

AccessLearn is an accessible, multi-modal AI learning companion powered by Google Gemini 2.5 Flash. It takes any educational passage (or textbook excerpt) and transforms it into a transparent 5-step pedagogical loop:

  1. Grounded Multi-Modal Explanations: Converts complex texts into one of three accessible modes without hallucination:
    • Simple: Plain language unpacking technical jargon.
    • Analogy: Vivid, intuitive everyday metaphors (e.g. high-fives and bathroom scales for Newton's 3rd Law).
    • Step-by-Step: Sequential cause-and-effect breakdowns. Also extracts key technical terminology with interactive definitions and supports ElevenLabs audio roadmaps.
  2. Targeted Conceptual Assessment: Synthesizes a focused comprehension question assessing underlying mechanics rather than shallow keyword recall.
  3. Semantic Evaluation & Evidence Extraction: Evaluates the learner's free-form natural language response without penalizing informal phrasing or unique perspectives.
  4. Respectful Misunderstanding Diagnosis: When a learner falters, it doesn't give a generic error. It identifies the Likely Conceptual Misunderstanding (e.g., confusing action-reaction pairs with static equilibrium), directly quotes the student's own words as evidence, and respectfully explains what the source material actually establishes.
  5. Adaptive Re-Explanation Loop & Safeguards:
    • Offers an alternative explanation and an intuitive counterexample (e.g., rocket propulsion) with a fresh retry question.
    • Includes a Full Worked Conceptual Solution modal safeguard to prevent frustration and ensure zero penalties for exploring reasoning.
    • Tracks session progress across all concepts in a persistent drawer.

How we built it

  • Full-Stack Framework: Next.js 16 (App Router) with React 19 and TypeScript for strict type safety and fast server-rendered accessibility.
  • Styling & Accessibility: Tailwind CSS v4 featuring high-contrast typography, accessible focus indicators (:focus-visible), aria landmarks, and warm, calming cream/violet design tokens (#FAF8F5, #5B45E0).
  • AI Core: Official @google/genai SDK orchestrating Gemini 2.5 Flash with strict JSON schemas (responseSchema) to enforce zero hallucinations and guaranteed semantic parsing.
  • Speech & Audio: Windows SAPI speech synthesis and architecture-ready ElevenLabs speech preview pipeline.
  • Testing & Verification: 8/8 comprehensive automated unit tests covering schema validation and pedagogical state transitions, alongside Playwright end-to-end automation for pristine headless recording.

Challenges we ran into

  • Preventing AI Hallucinations: Generic LLMs love to fabricate extraneous facts when explaining concepts. We solved this by implementing strict system prompts that ground every generated sentence directly in the user-provided passage.
  • Semantic Nuance in Student Answers: Distinguishing between an answer that is technically imprecise vs. conceptually wrong required delicate prompt engineering. We separated evaluation into a two-tiered system: Supportive Feedback (validating correct observations) and Misunderstanding Diagnosis (addressing the root conceptual gap).
  • Video & Visual Framing: Perfecting smooth camera motion across dynamic responsive cards without sticky-header clipping required custom Playwright easing scripts.

Accomplishments that we're proud of

  • A clean, transparent learning loop where the student is never penalized for making a mistake.
  • The Likely Misunderstanding diagnostic engine that cites exact student quotes and explains the underlying mental model gap.
  • 100% test pass rate with zero lint or build errors on Next.js 16 and Turbopack.
  • High-performance, low-latency live responses powered by Gemini 2.5 Flash.

What we learned

We gained deep insights into prompt calibration with structured output schemas in the new @google/genai SDK, accessible UX design patterns for neurodivergent learners, and how to turn diagnostic AI feedback into an encouraging, supportive learning experience.

What's next for AccessLearn

  • Full ElevenLabs Multi-Voice Streaming: Real-time streaming voice with word-level karaoke highlighting for auditory learners.
  • PDF & Textbook OCR Upload: Upload entire lecture slides and syllabus chapters to generate interactive study decks.
  • LTI Canvas/Blackboard Integration: Allowing educators to embed AccessLearn directly into university LMS courses.

Built With

Share this project:

Updates