CTO @ Zilix | UC Berkeley EECS
Software
Portfolio
I'm a software engineer based in El Dorado Hills, CA, and an EECS student at UC Berkeley. Right now I'm Founding Engineer and CTO at Zilix and a Junior Software Developer at Protelo, and most of my work lately is on AI agents: governing them, training them, and shipping products with them. Thanks for stopping by, and keep scrolling to see what I've been building!
Founding Engineer & CTO · Founders Inc. Off Season II · Jun 2026 – Present
Zilix is a desktop app that sits between AI coding agents (Claude Code, Codex, local models) and the developer's machine: every tool call is checked against policy before it runs, consequential actions wait for a human, and everything lands in a hash-chained, tamper-evident audit log. I designed the governance model, where an organization's signed policy can only take permissions away, never grant them. We've brought frontier AI coding into regulated enterprise codebases, reaching 30+ developers, including developers at Bank of America, Micron, and Mirdain Systems.
Junior Software Developer · Apr 2026 – Present
I shipped RightPhone to 1,000+ daily active users across enterprise customer orgs. It replaced 3 separate tools by putting RingCentral WebRTC and NetSuite CRM behind a single Next.js interface, with every call and SMS auto-logged to the CRM. I also built a real-time WebSocket signaling and session layer with JWT / OAuth 2.0 auth for concurrent multi-tenant sessions, and architected RightInsight, a multi-tenant analytics platform that ingests NetSuite saved-search data and routes it through Gemini 2.5, cutting reporting cycles from hours to minutes.
I started coding in middle school, drawing simple graphical interfaces with Python's turtle module. At Folsom Lake College I was President of FLC++, the coding club, growing it to 322 members, managing 5 officers, and running workshops, hackathons, and engineering projects like CouchMo. Two years and a 4.0 later, I transferred to UC Berkeley, where I'm working toward a B.S. in EECS, expected Dec 2028.
Python, Java, C++, Rust, JavaScript/TypeScript, SQL, HTML/CSS
PyTorch, TorchRL, Stable-Baselines3, TensorFlow, ONNX Runtime, OpenCV, Pandas, Jupyter
React, Next.js, Node.js, Electron, Express, Django, Flask, REST, WebSockets, WebRTC, JWT/OAuth 2.0
AWS, Docker, Vercel, Railway, GitHub Actions (CI/CD), ROS 2, Bitbucket, Git, PostgreSQL, MongoDB, MySQL
🏆 Best Club Project, FLC++ · Demo Video
An autonomous self-driving couch. The policy learns by behavioral cloning in a containerized Gazebo + ROS 2 simulator on rollouts from an Intelligent Driver Model expert, then gets fine-tuned with PPO via Stable-Baselines3 and deployed through ONNX Runtime to an ESP32-driven chassis.
Technologies: Python, PyTorch, Stable-Baselines3, ROS 2, Gazebo, Docker, ONNX Runtime, OpenCV
A PPO agent that beats Level 9 CPUs, the highest in-game difficulty, in Super Smash Bros. Melee despite frame-perfect inputs and an adversarial opponent. An ActorCritic MLP trains for 500k–1.5M frames on a custom Melee physics simulator, then the Fox policy is fine-tuned in real Dolphin via libmelee + Slippi against a trained Jigglypuff, with shaped reward terms for combos, edgeguards, and shield pressure. A FastAPI OpenEnv server launches Dolphin and exposes reset()/step() over HTTP.
Technologies: Python, PyTorch, TorchRL, PPO, libmelee, Slippi Dolphin, Meta OpenEnv
🏆 1st Place, Best use of Modulate's APIs · Autonomous Agents Hackathon (200 participants, AWS Builder Loft)
An agent that finds peak-engagement moments in a Twitch stream from chat velocity, keyword trends, and real-time emotion analysis, then clips and publishes them. FFmpeg extracts audio, Modulate's REST API returns transcripts and emotion labels, Reka writes the caption, and the Instagram Graph API posts it. Roughly 2 hours of manual editing becomes a ~30 second autonomous pipeline.
Technologies: Python, FFmpeg, Twitch API, Modulate, Reka, Instagram Graph API
A multi-agent courtroom simulation where AI agents argue using real-case information scraped with Tavily and a RAG-powered engine, benchmarked against each other on how well they cite legal precedent.
Technologies: React, LangGraph.js, Nebius Open LLMs, Tavily
A dashboard for monitoring EV charging stations inside a real-world Monte Carlo simulation, with AI agents automating maintenance to reduce downtime.
Technologies: TypeScript, Next.js, Groq, LangGraph.js
The room and its contents were modeled in Blender, then rendered and animated with Three.js. Try switching to dark mode to see the full experience!
Technologies: JavaScript, Blender, Three.js, HTML/CSS