Hierarchical Classifier
Dynamic taxonomies through vector embeddings
Demo
Overview
Research project at Utah Tech University exploring hierarchical classification without retraining. I designed embeddings using Python encoding both semantic meaning and a node's position within arbitrary hierarchies—bridging meaning and structure for flexible taxonomy generation.
Role: AI Research Engineer
Technical Stack
- •Python
- •FastAPI
- •PostgreSQL
- •OpenAI Embeddings
- •AWS (EB, RDS)
Notable Features & Challenges
- 1Multi‑level clustering algorithm for arbitrary hierarchies
- 2Embedding pipeline encoding semantics + structure
- 3Evaluation framework for similarity precision vs. latency
- 4Optimized inference to 400 ms on > 50 req/s load
Impact
Reached 90 % accuracy with dynamic hierarchies; processed 50+ req/s in sub‑400 ms latency. Performance validated through partner Zonos project.
Reflection
Taxonomy design is an ancient philosophical problem repackaged as data science. Encoding 'what belongs with what' required as much conceptual rigor as coding skill. This work fused my philosophical interest in categories with applied AI engineering.