Selected Work
A trajectory from constructivist AI learning environments, to retrieval evaluation, to long-term supportive AI memory.
This portfolio represents a deliberate progression toward designing AI systems that guide human reflection without removing agency. By starting with grounded educational simulations and moving into the technical underpinnings of memory retrieval, the ultimate goal is to build long-term supportive AI architectures.
MetaClues: Constructivist Learning in Metaverse
(IEEE MetaCom 2025)
MetaClues explores how immersive, constructivist learning environments can leverage AI to provide contextual scaffolding without overriding the learner's discovery process. The project investigates the delicate balance between helpful intervention and maintaining cognitive friction necessary for learning.
CONTRIBUTION
Designed novel interaction mechanics for spatial AI agents that offer hints based on gaze and interaction history.
RESEARCH RELEVANCE
Establishes baseline methodologies for evaluating user agency in highly structured, AI-mediated environments.
NBChat: Year-Long Prototype for Long-Term Memory
NBChat serves as a longitudinal study platform for investigating how AI systems can develop and maintain a coherent understanding of a user over extended periods. It focuses on the architectural challenges of episodic memory consolidation and retrieval in conversational interfaces.
ARCHITECTURE
- FastAPI backend for asynchronous inference
- MongoDB for hybrid semantic/lexical search
- React Native mobile client
TRANSITION REFLECTION
Bridged the gap between theoretical memory models and practical engineering constraints for real-time interaction.
RAG & Embedding Fine-Tuning
Providing the NLP technical grounding required for robust memory systems. This project involved fine-tuning embedding models on domain-specific conversational datasets to improve retrieval accuracy in Retrieval-Augmented Generation pipelines.
COMPONENTS
Contrastive learning techniques applied to Sentence-Transformers, vector database optimization, and custom reranking heuristics.
RESULTS
Achieved a 24% improvement in top-k retrieval accuracy for ambiguous conversational queries compared to baseline models.