TALIA
An educational AI assistant that explains machine learning concepts the way she wishes they'd been explained to her — visually, and from first principles.
Problem
Most ML explainers are either too shallow to be useful or too dense to be approachable — a gap she noticed firsthand while moving from design into ML herself.
Approach
Built a retrieval-augmented assistant grounded in a curated set of primary sources and her own written explainers, paired with generated visualizations so answers show the underlying structure rather than just describing it in words.
Architecture
Curated knowledge base → embedding index → retrieval-augmented generation with citation tracking → on-demand diagram generation for concepts with clear visual structure.
Results
- Used as a companion tool alongside her educational content and talks.
- Positive feedback specifically on the paired visual explanations for embedding and attention concepts.
Future work
Expanding the knowledge base to cover scientific ML topics and adding interactive, adjustable diagrams for concepts like attention and positional encoding.