StAnify: Multi-Agent Visual Storytelling & AI Tutoring
Prototype
Agent flow — click any component (or focus the graph and use arrow keys) to see what data moves through it
?
Click a component, or use the arrow keys, to inspect it.
Architecture and numbers from the thesis: the CrewAI pipeline with its Qdrant RAG store and regeneration loops, and the blinded A/B results against GPT-4 and GPT-5.
Highlights
- System Architecture: Conceptualized, designed, and constructed a distributed multi-agent pedagogical framework utilizing CrewAI to deliver age-appropriate, curriculum-aligned educational modules for early childhood education (ages 5–9).
- Agent Coordination: Orchestrated an interconnected network of specialized agents managing unique scopes (analogy creation, narrative generation, structural vocabulary scaffolding, text chunking, automated criteria evaluation, Stable Diffusion XL image generation, and multi-modal vision feedback critique) handled via stream-gated Pydantic, YAML/JSON schemas, and Streamlit frontends.
- RAG Framework: Integrated an advanced Retrieval-Augmented Generation pipeline using Qdrant vector databases to guarantee semantic consistency and zero hallucinations during content creation.
- Advanced Code Optimization: Leveraged state-of-the-art developer systems (Claude Sonnet, GPT-4o, Copilot, Cursor AI) to write clean, maintainable python codebases containing custom error boundaries and asynchronous processing constraints.
- Blinded A/B Testing: Conducted a rigorous, blinded comparative study involving 30 domain evaluators assessing generated output across Mathematics, Biology, and Science modules against standard baseline GPT-4 and GPT-5 generations.
- Empirical Findings: Discovered that StAnify was significantly preferred in 65% of all target test cases, scoring profound measurable improvements over GPT-4 in aesthetic visual clarity (+1.59) and student engagement (+1.51) while matching the absolute conceptual accuracy of GPT-5.
- Efficiency Metrics: Compressed traditional manual classroom concept preparation times down from 50–80 minutes to ~3 minutes flat (representing a 17x–25x speedup) at a marginal production cost of ~$0.15 per conceptual package.
Technologies
- CrewAI
- Qdrant
- Stable Diffusion XL
- Streamlit
- Pydantic