Featured Work
SJC Airport AI Concierge
Production RAG-powered customer service chatbot for San Jose Mineta International Airport, live at flysanjose.com.
A tool-augmented RAG pipeline built with LangChain, OpenAI embeddings, and Qdrant, grounding responses across 13,000+ chunked documents pulled from 667 pages of the SJC website.
A Redis-based semantic cache with hash-value change detection and automated refresh, cutting inference costs by 25% and improving end-to-end response latency by 66%.
A model selection evaluation suite using an LLM-as-Judge approach across 200+ benchmark Q&A pairs, improving answer accuracy from 75% to 98% and reducing response latency from 5s to under 3s. Deployed serverless on Google Cloud Functions and Firebase with retry logic and graceful failure handling.
View live at flysanjose.com— chatbot is in the bottom right corner
Personal Projects
PrepAgent
AI Interview Research AgentJune 2026 — Present
An MCP-native multi-agent system using a LangGraph supervisor graph with 4 specialized agents, integrating Gmail, Google Calendar, and Google Drive via Google OAuth to autonomously detect interview invites and trigger a personalized research pipeline without manual user input. Features a RAG pipeline with Pinecone, Cohere Rerank, and OpenAI embeddings to retrieve and rerank context across heterogeneous sources, with a RAGAS eval suite tracking answer relevance, faithfulness, and context precision. Targets sub-2-minute end-to-end generation for research that currently takes 45 minutes of manual review.
Optimal
AI Productivity AgentJune 2025 — Jan. 2026
A supervisor-orchestrated multi-agent system with structured agent handoffs and human-in-the-loop approval gates, so agents propose actions for user review before anything gets written. Built offline evaluation benchmarks and regression suites to measure agent task completion and output quality, cutting prompt regression incidents by 40% across 3 model update cycles. Full-stack with React Native (Expo), Spring Boot, and a Supabase PostgreSQL schema with 17+ entities. Distributed via TestFlight for real-world testing.
Quizzler
AI Quiz GeneratorMar. 2025, Updated Apr. 2026
A 5-stage RAG pipeline instrumented with LangSmith tracing across query expansion, embedding, reranking, and generation, capturing token usage, Cohere rescores, and per-span latency to surface retrieval quality and cost regressions. Features parallel upload processing to chunk and embed user documents into an ephemeral server-side vector store while concurrently generating structured MCQ quizzes, delivering grounded answers with glass-box citations from user-uploaded PDFs and text.
Skills