Leading research and development on next-generation knowledge graph platforms and agentic AI systems at the AI Institute.

Leading the effort in developing a Neuro-symbolic AI approach involving small language models and knowledge graphs to ingest medical device documentation, process guidelines, and patent submissions to accelerate and automate medical device R&D.

• Directed collaborative data science discovery projects with UC Berkeley to develop a medical device knowledge graph to support complex multi-hop reasoning and information retrieval with provenance from large-scale FDA 510(k) datasets
• Engineered a Multi-Strategy (Document/Graph)-RAG with LLM-as-a-Judge approach to improve response traceability
• Designed a KG-LLM response fidelity score to evaluate and audit the lineage of agents outputs and reduce hallucinations in medical device QA

Pioneered knowledge graph initiatives to transform sales engagement data into actionable intelligence for enterprise sales teams.

• Designed the Sales Engagement Ontology to unify semantic standards between data producers and consumers
• Pioneered the Sales Engagement Knowledge Graph (SEG) to surface the people's information in sales activities from 4.7M engagement logs; improved the sales-rep-to-lead connections throughput by 20% (~64K newly discovered people entities)
• Modeled the non-sequential sales processes with multi-label temporal graphs; transitioning from a linear "stages" view to a multi-dimensional "activity" view to improve context and next best actions

Developed context-enriched deep learning models to align knowledge from >200 heterogeneous sources for the UMLS Metathesaurus; streamlined the maintenance effort with humans in the loop.

• Benchmarked performance with various Knowledge Graph Embedding techniques and Siamese Network architectures to scale biomedical vocabulary alignment across 172 millions term pairs; achieved a 5.0% increase in precision (>94% F1) and a 50% reduction in false positive rates compared to lexical baselines
• Optimized large-scale training pipelines using Keras, TensorFlow, and Slurm on high-performance computing clusters, utilizing over 2,700 GPU hours for models training and validation