Devpost submission text — EchoScholar (ML Empowerment Build Challenge 3.0)

Problem Statement

Students study from lecture videos, textbooks and slides scattered across platforms — and generic AI chatbots answer from anywhere except the course material, or invent answers when the material doesn't cover a question.

For students who learn by listening — blind and low-vision students, students with dyslexia or reading fatigue, commuters, and anyone whose best focus happens away from a desk — the problem is worse: almost every study tool is screen-first. Text-heavy interfaces, silent quiz apps and silent feedback lock out the students who most need a trustworthy study companion.

Solution Overview

EchoScholar is an audio-first, source-grounded study companion. Upload your own course material (PDF textbooks, PPTX slide decks, lecture transcripts, notes), and EchoScholar:

1. Builds a source-cited knowledge base — every passage keeps its exact origin (page / slide / timestamp).
2. Answers questions only from that material, with citations — and declines questions the material doesn't cover instead of hallucinating.
3. Speaks: every answer is read aloud automatically; quiz questions can be listened to; feedback is spoken; and one button plays a personalized audio study brief built from the learner model.
4. Tracks mastery per topic with Bayesian Knowledge Tracing, so the audio brief tells the student exactly what to review next.

The audio path is an equal path, not a fallback: the same content stays available as text with one-click source verification for sighted users and screen readers.

Key Features

• Multimodal ingestion — PDF / PPTX / TXT / MD / VTT / SRT with zero manual preprocessing.
• Grounded tutor with refusal — answers cite [unit_id]; off-material questions return an explicit refusal; a citation click opens the exact excerpt.
• 🔊 Spoken answers — neural TTS (edge-tts), content-hash cached so repeats are instant.
• 🔊 Audio quiz mode — MCQ / short / numerical questions, each independently verified by a second model pass (mismatched keys are dropped or repaired), with repetition blocked across sessions.
• 🔊 Spoken feedback — model answer + explanation read aloud after grading.
• 🔊 Personalized audio study brief — weak topics, mastery %, and next steps, spoken.
• Learner model — per-topic BKT mastery + a 7-day study plan weighted by forgetting curves.
• Accessible UI — large type, high contrast, keyboard-first (Ctrl+Enter to ask), ARIA-labelled controls.

Target Users

• Blind and low-vision students who need study tools that work without a screen in front of them.
• Students with dyslexia or reading fatigue who absorb material better by listening.
• Commuters and auditory learners who want hands-free, eyes-free revision.
• Any student who wants a tutor they can trust: cited answers, honest refusals, no hallucination.

Social Impact Statement

Education is the most scalable form of empowerment, and accessibility is where it most often fails. Most "AI tutor" tools are built for sighted, screen-attached users; EchoScholar was designed from the audio side first, so the students most often left behind get the better experience, not a degraded one. Source grounding adds trust: a study companion that refuses to invent is safe for unsupervised learning — and teaching students that "I don't know" is an acceptable answer is itself a lesson worth scaling.

How We Built It

• Python + FastAPI backend; single-page vanilla-JS UI (no build step).
• Retrieval: scikit-learn TF-IDF (local, any language) + optional LLM re-ranking.
• Grounding: retrieval-scoped prompts, mandatory citations, explicit refusal token.
• Quiz pipeline: generate → structural validation → independent solve-verification → novelty hashing.
• Audio: edge-tts neural voices, disk-cached by content hash; multi-language voices via config.
• Learner model: Bayesian Knowledge Tracing with question-type aware priors.
• LLM: any OpenAI-compatible endpoint.

Challenges

• Keeping an LLM honest about refusals needed a strict prompt contract plus UI-level flags.
• Independent question verification caught real answer-key errors that generation alone missed.
• Audio caching made the audio-first experience fast (repeat synthesis: instant).

What's Next

• Speech-to-text for raw lecture videos; OCR for diagrams.
• Voice input so the whole loop is eyes-free.
• IRT/DKT extensions of the learner model; multi-language tutoring.

Built with

python, fastapi, edge-tts, scikit-learn, bayesian-knowledge-tracing, llm, rag, accessibility

Built With

• accessibility
• edge-tts
• fastapi
• llm
• python
• rag
• scikit-learn