Research
Similarity-based retrieval has become a default design choice in LLM systems, but this assumption breaks down in high-stakes decision environments. In maternal and clinical contexts, retrieving semantically similar text does not guarantee actionable, safe, or context-valid guidance.
My research addresses this gap by combining utility-driven retrieval objectives with neuro-symbolic reasoning and explicit safety constraints. The goal is to build systems that can justify retrieval choices, reduce hallucinations, and maintain reliability under real-world uncertainty.
A research system that replaces purely similarity-based retrieval with utility-driven evidence selection. It integrates graph-based retrieval reasoning and symbolic safety constraints to improve output validity in maternal health decision support.
Developed a retrieval-augmented clinical assistant using Flask and LLaMA, then performed grounding and hallucination analysis under varying retrieval conditions. The work identifies failure pathways in medically sensitive prompts and proposes mitigation strategies.
Designed and implemented a web-based education system to support maternal and neonatal health literacy with structured content pathways and interactive guidance. The project informs user-centered evaluation for future intelligent assistants.
Built a machine learning classification system for early risk indication and interpretation of relevant patient features. The project emphasized robust preprocessing and transparent performance evaluation.
Prototyped an adaptive tutoring system that personalized hints and sequence progression based on learner responses. The system explored lightweight decision logic for low-resource deployment settings.