01. Problem Context & Data Fragmentation
At a major research institution like UIUC, indexing and cataloging faculty research is an operational challenge. Across dozens of departments and thousands of researchers, research profiles existed as unstructured, heterogeneous text descriptions with conflicting departmental vocabulary.
Taxonomy Mapping & Semantic Ambiguity
Because academic research is increasingly interdisciplinary, keyword matching failed to resolve complex research domains. Free-form text required an automated semantic extraction pipeline that could enforce strict hierarchical taxonomy constraints.
LangGraph
Stateful cyclic graph
Pydantic
Strict schema validation
Gemini Flash
Low-latency API inference
02. Pipeline Architecture
To solve this, I engineered an automated agentic pipeline that ingests unstructured natural language and guarantees verified, schema-compliant research classifications.
Description Normalization
Normalizes raw faculty profile texts, stripping noise and formatting artifacts before downstream processing.
Tiered Classification
Moves sequentially through a High-Level Field Classifier into a granular Subfield Classifier, triggering a fallback Subfield Proposer for edge cases.
Strict Validation
Forces classification payloads through Pydantic models to guarantee exact alignment with official university taxonomy trees.
03. Architectural Decisions & Trade-Offs
Framework: LangGraph vs. Linear Chains
Selected LangGraph over linear chains to enable stateful, cyclic graphs with conditional retry loops, ensuring edge-case classifications could be iteratively refined.
Inference Efficiency: Gemini Flash
Benchmarked model response times and throughput, selecting Gemini Flash for the primary classification nodes to minimize latency and manage API resource ceilings.
Human-in-the-Loop Curator
Built a human-in-the-loop review queue for out-of-vocabulary proposals, preventing taxonomy degradation while surfacing novel interdisciplinary topics.
04. System Deployment & Impact
Taxonomy Standardization
Delivered an automated categorization engine that maps disparate academic descriptions into a unified taxonomy, replacing fragmented departmental labels with a consistent computational schema.
Programmatic Discovery
Enabled structured semantic queries across hundreds of faculty profiles, supporting programmatic indexing and interdisciplinary research discovery across campus.