Built a multi-agent investing model that runs five specialized AI agents in parallel for every stock — value, growth and momentum, macro regime, quality/risk, and sentiment. Each agent computes quantitative scores, pulls grounded articles via Perplexity, and hands everything to OpenAI for a written rationale. Scores blend under client-aware weights and pass an IPS gate before becoming a portfolio. A live Streamlit demo is embedded on the AI Agents page.

Implemented a rules and retrieval lookup system over roughly 200 pages of FIRST Robotics competition manual content. The system returns cited excerpts and linked images so team members can quickly reference rules during robot design, inspection preparation, and competition planning.

Applied learnings from deployed client systems to a research project on adaptive retrieval and response-quality evaluation across heterogeneous datasets. The work studies how query-specific choices in retrieval depth, chunking, model selection, and generation settings affect answer quality, latency, and cost.

Built a RAG-based student chatbot for a university business school using over 7 years of instructional letters and course archives. The assistant was optimized for source-grounded responses and consistency with the professor's teaching voice.

Led end-to-end development of a digital sommelier for a wine publication using a 22+ year archive containing over 3 million words. The system implements quote extraction, page-level citation, and evidence packaging so answers remain grounded in published articles.