Inspiration
AI study tools sound confident even when they're wrong. Students can't tell which claims to trust. Athena's founding rule: no answer without a source you can check. The adaptive layer (Elo + spaced repetition) exists because static quizzes waste time — Athena meets you at your level on every topic.
What it does
Upload your study material — text, Markdown, PDFs, and diagrams with a short caption describing what they show — and Athena chunks it into a pure-Python BM25 search index. Ask anything in the Study tab: Athena retrieves the most relevant passages and answers using NVIDIA's Nemotron 3 Super — but it may only answer from those retrieved passages. Every factual claim carries an inline citation ([1], [2], …), and the exact source text sits in a Sources panel beside the answer. If your notes don't cover the question, Athena says so instead of inventing facts. Then test yourself in the Quiz tab: Athena generates questions grounded in your notes — multiple-choice and fill-in-the-blank, never generic trivia. Each topic carries a skill rating modeled on chess Elo: answer correctly and the next question gets harder; struggle, and it eases off and explains with citations. Wrong answers return later through Leitner spaced repetition, right when you're about to forget them. The Progress tab turns all of this into a mastery map — per-topic ratings, questions answered, reviews due — so you always know what to study next.
How we built it
Backend: FastAPI + SQLite. Pure-Python BM25 retrieval (no embedding service needed), per-topic Elo ratings, Leitner boxes. LLM: NVIDIA Nemotron 3 Super via an OpenAI-compatible inference endpoint, used for two grounded tasks only — cited Q&A and MCQ generation. A template-question fallback keeps quizzes working if the model is unreachable (the UI labels it "template mode"). Frontend: vanilla JS single page, restrained light/slate/indigo design. Images are registered with user-provided captions; Athena does not claim automatic visual understanding — the caption text is what gets indexed and quizzed.
Challenges we ran into
Keeping the model honest: the Q&A prompt constrains the model to retrieved passages and requires citations for every factual claim; we validate citation markers and fall back to a retrieval-only answer when the model is unavailable. Flaky inference endpoints: transient 503s from the hosted model forced us to build graceful degradation (template quiz mode, retrieval-only answers) instead of hard failures.
Accomplishments that we're proud of
A genuinely source-grounded study loop — citations are validated, not decorative. An adaptive engine (Elo + Leitner) that personalizes difficulty per topic. It runs end-to-end from one command: bash run.sh.
What we learned
Grounding is a systems problem, not just a prompt problem: retrieval quality, citation validation, and honest fallbacks matter as much as the model. And adaptive testing only works if the difficulty signal (Elo) updates on every answer.
What's next for Athena
Exportable study packs and shareable quizzes. More question types (short answer with rubric grading). Optional local-model mode for fully offline study.
Log in or sign up for Devpost to join the conversation.