At AXION Ray, I built a RAG pipeline that maps risk narratives to regulatory Annex codes, replacing existing multi-day manual tagging. It achieved 88% Recall@5 across 35K+ narratives.
I also developed an algorithmic mapping solution for disconnected datasets without primary keys, improving labeled-match accuracy from 61% to 82%. I operationalized GPT-4.1 entity tagging and an LLM classification pipeline with human-in-the-loop expert evaluation.
At Ushur Inc., I developed XGBoost classifiers to identify high-risk churn accounts for enterprise healthcare and insurance clients, achieving 0.86 ROC-AUC on held-out data. I also supported analytics for 20+ enterprise clients and helped reduce data-model build times by 85%.
On a medical device complaint classification project, I fine-tuned Llama 3.1 8B using 4-bit QLoRA, improving macro F1 from 0.62 to 0.88. I also built an agentic database operations assistant that translates natural-language requests into SQL CRUD operations.
