About
I'm an AI engineer finishing my M.S. in Artificial Intelligence at San José State University (Dec 2026). As a Graduate Research Assistant, I'm building a RAG-based literature classification system that combines ChromaDB hybrid retrieval with LLM judging, validated against hand-labeled ground truth.
Before grad school I spent 3.5 years as a software engineer at King Children, a 3D-eyewear startup. I was the founding engineer on the B2B partnership portal that moved the company from direct-to-consumer to B2B, and I owned production systems across Next.js, React/GraphQL, Python/Tornado and PostgreSQL.
I care about LLM systems that are measured, not vibe-checked: evals, tracing and ground-truth labels. My multi-agent RAG pipeline reached 48× higher entity F1 and cut hallucination from 8.0% to 5.3%. I'm looking for a full-time role starting January 2027 where I can ship AI products end to end.
LLM systems, measured
RAG pipelines, fine-tuned models and applied ML, each with evals and numbers behind it.
48× entity F1 · hallucination 8.0% → 5.3%
Three-agent LangGraph pipeline (Scout → Mapper → Critic) over FAISS dense retrieval on a 195-document corpus. Benchmarked four LLMs (GPT, Gemini, Llama, Nemotron) on 55 hand-labeled queries against a zero-shot baseline. Served via FastAPI with a multi-model comparison UI and per-agent Langfuse tracing.
Medical NLI with TinyClinicalBERT
Fine-tuned TinyClinicalBERT on the MedNLI dataset with PyTorch and Hugging Face, using custom training loops, evaluation metrics and checkpointing to classify entailment between clinical statement pairs.
AI-Enhanced WiFi Sensing for Security
Camera-free security sensing using RuView DensePose and ST-GCN models over WiFi signal data from a network of ESP32-S3 sensor nodes.
Ask-Me-Anything RAG Chatbot (this site)
Try it: bottom-right corner
The chat assistant on this page. Site content and my resume are embedded with Gemini embeddings, retrieved by cosine similarity (top-k = 4), and answered by Gemini 2.5 Flash with grounding-only instructions. Built with LlamaIndex documents and a Next.js API route.
More projects
- Predictive Modeling of Flight Delays at SFO2025
End-to-end classification pipeline over 135,000+ SFO departure records to predict delays past 15 minutes; XGBoost led on F1 after feature engineering, importance analysis and leakage controls.
4-person offline-RL study. I designed the controlled-shift MuJoCo wrappers (gravity, friction, observation noise, reward) and the evaluation pipeline behind 1,536 evaluations on SJSU HPC; a triple-critic ensemble cut gravity AUDC by 14% on Hopper.
- AI-Based Enterprise Help Desk2025
Four-person team build of an IT help-desk platform with LLM-based ticket categorization (OpenAI API), Auth0 SSO role-based access, DynamoDB, and a React front end shipped through AWS CodePipeline.
Experience
- Implementing a candidate-paper relevance classification system with ChromaDB hybrid retrieval and LLM-based judging, validated through a stratified, hand-labeled ground-truth workflow.
- Compiled a 70,000-paper transportation-research corpus with a Scopus API enrichment pipeline, organized under a 370-term taxonomy across 4 research domains for downstream retrieval.
- Founding engineer on a new B2B partnership portal (Next.js 14, TypeScript, GraphQL) that pivoted the company from direct-to-consumer to B2B: PRD to production MVP in 10 weeks across ~8 routes and 40+ components, leading to 2 new partnerships after launch at Vision Expo East 2024.
- Architected a Notion-based CRM integration with a webhook that auto-populates design-task deadlines across 3 databases, plus a process-tracking dashboard optimized from an N+1 bottleneck to a single batched query.
- Maintained the Alembic migration history on a live production PostgreSQL schema: ~20 migrations, each staging-verified and peer-reviewed, with no production incidents.
- Authored 5+ product requirement documents with cross-functional stakeholders, B2B partners and customers to scope new features.
- Built and owned the internal order-management console (React, Redux, Apollo/GraphQL) used daily by fulfillment: a paginated, multi-facet-filterable orders table with bulk status, material and date editing.
- Built a data-automation tool pulling design images from AWS S3 and measurements from PostgreSQL into HTML comparison grids, cutting a designer's weekly effort from 2 hours to 10 minutes.
- Built the order "on-hold" workflow end to end: schema migrations, backend status logic, a React staff UI, and a scheduled Slack summary job.
- Shipped a discount-code management and batch-generation engine (React, Python/Tornado) with multi-currency support for marketing and partnerships.
- Built a reusable UIKit/SwiftUI bridging utility so a seven-year-old iOS app could adopt SwiftUI incrementally, and delivered Firebase push notifications end to end across the Tornado backend and Swift client.
Education
- Louie Barozzi Scholarship, awarded annually to 1 international grad student for academic excellence and campus leadership
- Coursework: Machine Learning, Deep Learning, Data Mining, AI & Data Engineering, Enterprise Software Platforms
- Treasurer, NextGen | German Connect · International Peer Mentor
Skills
Languages
AI & LLM
Backend & Data
Frontend
Cloud & Practices
AI Developer Tools
Let's talk
Hiring for an AI, ML or full-stack engineering role? I'm available full-time from January 2027 and usually reply within a day.