Case Study: LectureLens — AI Study Planner

Students juggle heterogeneous materials (slides, lecture notes, textbook PDFs) that vary in density and formatting. Translating this into a sustainable plan with spaced repetition and concept reinforcement is tedious and error-prone. Repetition without structure leads to plateaued retention and wasted time.

Product Flow

1. Import course material (notes, PDFs, slides)
2. Generate summaries and flashcards
3. Ask follow-up questions in Q&A
4. Create a study schedule with milestones

Architecture

• Next.js (App Router) for UI and server actions
• Supabase for authentication, Postgres, and secure storage
• TypeScript for correctness across the stack

AI Processing Pipeline

1. Extraction & Segmentation: Uploaded PDF/notes are parsed into logical blocks (headings, paragraphs, enumerations) to preserve hierarchy.
2. Chunk Formation: Blocks grouped under a token threshold to keep prompt context focused and predictable in cost.
3. Per-Chunk Summarization: Summaries generated with consistency directives (terminology normalization) to reduce drift across sections.
4. Global Consolidation: Merge overlapping concepts; eliminate redundant glossary-style lines.
5. Flashcard Generation: Pattern-based prompts classify facts vs processes vs edge cases; dedupe via similarity scoring.
6. Interactive Q&A: User questions map to most relevant blocks (future: vector search) before response synthesis.
7. Schedule Synthesis: Summary difficulty + card density drive milestone allocation (intro → drill → review cycles).

Data Model

• documents: id, title, original_filename, processing_state
• blocks: id, document_id, ordinal, text, type (heading|paragraph|list)
• summaries: block_ids[], text, level (chunk|section|global)
• cards: question, answer, source_block_ids[], difficulty, next_review_at
• schedules: milestone_index, start_at, focus (new|review), target_card_ids[]

Key Constraints

• Token Limits: Chunk sizing prevents model context overflow and outlier latency.
• Terminology Drift: Consolidation normalizes synonyms to avoid mismatched flashcards.
• Redundancy: Similar blocks (slide bullets vs notes) filtered before card generation.
• Fair Scheduling: Milestones weighted by conceptual density, not raw character count.

• Noisy PDFs: Hyphenated line breaks and multi-column layouts required normalization heuristics to prevent fragmented prompts.
• Duplicate Concepts: Overlapping slide + note content inflated card counts; similarity checks trimmed 20–30% redundant candidates.
• Prompt Consistency: Early versions produced variable answer formats; adding explicit output schemas (lists, definitions) improved dedupe.
• Scheduling Bias: Long early chapters dominated milestone allocation until normalized by concept density metrics.
• Auth & Isolation: Supabase row-level security ensures users only access their own documents and derived artifacts.

• Vector embeddings (pgvector) for semantic retrieval in Q&A.
• Adaptive spaced repetition using historical success rate per card.
• Calendar (ICS) export & sync.
• Collaborative study sets with shared progress overlays.