Curriculum Analyzer

Curriculum Analyzer reads a university course handout and builds a knowledge graph of what the course teaches. Exam questions are then mapped onto that graph, so faculty can see which topics a paper covers and which it misses. Built as a team of four; I worked on building the knowledge graph and extracting topics from it.

How It Works

• Handout to Graph: The handout is parsed with pymupdf, which keeps the math glyphs that pdfplumber drops. An LLM extracts a Neo4j graph of Course → Topic → Concept with learning objectives, and a second pass adds REQUIRES and RELATED_TO links between topics.
• Hybrid Retrieval: Every searchable node is indexed twice, as Neo4j fulltext and as vector embeddings. The two rankings are fused with reciprocal rank fusion, weighting the dense side 1.5×.
• Question to Topics: Retrieval narrows the candidates, then a small local model picks the topics a question actually tests. Questions that fall outside the syllabus are flagged rather than forced onto a topic.
• Figures: Pages with circuit diagrams or plots are detected by counting Bézier curves in the PDF (0 on prose pages, 178 to 525 on circuit pages) and escalated to a vision model.
• Swappable Models: Graph extraction, topic selection and embeddings are separate roles, each with its own provider. Everything runs locally on Ollama (qwen3:4b, mxbai-embed-large), with Gemini as an option for graph extraction.

Results

• 96% of questions have a correct topic in the top three.
• 1.0000 mean Jaccard similarity across 15 repeat runs, so the same paper always maps the same way.
• 9/10 topic-selection accuracy with the 4B-parameter qwen3 model.
• The CS F111 handout produces 13 topics, 79 concepts and 48 learning objectives.

Tooling

A Streamlit app (Build, Inspect and Ask tabs) and a CLI (build, index, ask), with Langfuse tracing, structured logging, pytest, ruff and ty.