StudyAI

Students struggle to convert raw lecture capture—live notes, voice memos, and unstructured material—into review-ready summaries and formative quizzes fast enough to reinforce learning while content is still fresh.

Close the gap between live lecture delivery and effective self-review by converting unstructured in-class input into retention-ready summaries and assessments on a mobile-first, latency-sensitive path.

• Mobile-first UX with intermittent campus connectivity—offline access required for post-lecture review
• Inference latency must stay low enough for interactive study sessions during and after class
• Quiz quality depends on summarization fidelity from noisy, real-time speech and note input

• Built in-class live note capture with speech-to-text ingestion, streaming lecture content into the processing pipeline without disrupting session flow.
• Deployed Meta BART (`facebook/bart-large-cnn`) for abstractive summarization and quiz-item generation—producing study briefs and formative assessment prompts grounded in captured material.
• Architected an offline-first SwiftUI client with Firestore + Core Data dual persistence and a FastAPI layer targeting p95 inference latency under 400ms.

• Selected Meta BART (`facebook/bart-large-cnn`) for bounded seq2seq summarization and text generation instead of open-ended LLM calls—better cost and latency control for structured study outputs
• Kept live capture lightweight on-device while delegating summarization and quiz synthesis to the backend inference path
• Used dual-layer persistence (Firestore + Core Data) so students can review generated material without continuous network access

• Offline-first sync for note and quiz access across intermittent campus connectivity
• Optimized inference paths with p95 response times under 400ms for interactive study workflows

Built an iOS study companion prototype that turns in-class live notes and voice capture into BART-powered summaries and auto-generated quizzes—structured review material from raw lecture input in a single mobile workflow.

• Not shipped; iOS prototype only—no App Store release
• Quiz quality varies with speech recognition accuracy and lecture audio conditions
• Model scope limited to summarization and generation—no multimodal or long-document RAG pipeline

• Add retrieval grounding for course materials (syllabi, slides) to improve quiz relevance
• Expand evaluation harnesses for summary and quiz quality across subject domains

• SwiftUI
• FastAPI
• BART (facebook/bart-large-cnn)
• Firebase
• Firestore
• Core Data
• Speech-to-text