SnapLearn

Inspiration

Students photograph textbook pages and lecture slides every day — then those photos die in the camera roll. We wanted to turn any study photo into an active learning session: not just "AI explains the image," but a full loop of explain → quiz → diagnose your misconceptions → re-explain.

What it does

SnapLearn is a multimodal AI study companion. Upload a photo of any educational material (textbook diagram, whiteboard, handwritten notes, lecture slide) and it:

1. Explains the concept with every claim cited to its source image ([Source: image-N]) — provenance you can check.
2. Quizzes you with auto-generated questions tagged by topic + difficulty (easy/medium/hard), never repeating a question.
3. Diagnoses misconceptions — answer wrong and the AI pinpoints exactly what you misunderstood and re-explains that point, with citations.
4. Answers follow-ups grounded in your uploads, and explicitly flags questions that go beyond the scope of your materials.

How we built it

• Backend: FastAPI (Python) with 4 endpoints — /api/analyze (image → explanation + quiz), /api/ask (grounded Q&A + scope guard), /api/quiz (more questions, de-duplicated), /api/grade (grading + misconception diagnosis).
• Vision-language model: NVIDIA NIM meta/llama-3.2-90b-vision-instruct via OpenAI-compatible API. Prompts enforce source citations on every factual sentence, scope limits, and structured quiz output.
• Frontend: single-file vanilla HTML/CSS/JS — drag-drop upload, quiz cards with difficulty badges, misconception feedback panels, chat-style Q&A. No build step, no accounts, no database (deliberately cut for scope).

Challenges we ran into

• The 90B vision model was slow (60–95s/call) but followed citation/quiz formatting; the faster 11B ignored formatting — we kept 90B with a 240s timeout.
• Getting the model to emit [Source: image-N] reliably took prompt hardening (few-shot examples + "missing citation = wrong answer" + temperature 0.2).
• 4.5 hours from idea to submission forced ruthless cuts: voice, accounts, history, mobile — all dropped to protect a working demo.

Accomplishments that we're proud of

• Real end-to-end multimodal loop verified live: upload → cited explanation → quiz → misconception diagnosis → re-explanation.
• Source-grounded answers with explicit out-of-scope flagging, matching Track D's provenance requirements.
• Honest engineering: mock mode is clearly labeled [MOCK] everywhere it appears.

What we learned

Multimodal learning apps live or die on trust: citations and scope honesty matter more than feature count. And a judge-friendly demo beats a 10-feature backlog.

What's next for SnapLearn

Multi-image knowledge bases with cross-image citations, spaced-repetition scheduling for weak topics, and lecture-video ingestion with timestamp provenance.

AI coding assistants (Muse/Claude) used throughout — full breakdown in README 'AI Assistance Disclosure'.

Built With

• css
• fastapi
• html
• javascript
• nvidia-nim
• python