Johnny Zhong builds
$ triage --watch feedback ▸ #4182 "preview freezes on export" route → bug, 0.92 (like #3907) logs → render thread stall ticket → filed, owner tagged reply → drafted for human check ✓ done in 3m 41s (was ~2 days)
Three software engineering internships, each closer to the place where models meet real users and real data.
Software Engineer Intern, Intelligent Creation
Software Engineer Intern, humanoid robotics
Software Engineer Intern
Things I build on nights, weekends and hackathons. Most of them are agents, or tools for working alongside agents.
A cross-platform desktop companion that turns live coding-agent events into tasks, rewards and approvals you can act on. I lead a team of three engineers and own the native macOS event bridges, the client state machines and the monorepo with its CI and release automation.
Rust, Tauri, Svelte, TypeScript
A real-time meeting display for people who can’t rely on hearing the room. It listens to the call, warns you before the conversation turns to you, and when you’re asked something it shows the question, the last few lines and a sentence you can say. In real meetings, people say your name in only 0.4% of lines, so Cue doesn’t wait for it.
TypeScript, Electron, native macOS audio capture, LLM judge
Paste a GitHub repository and fly through it as a glowing 3D city. Buildings are files, districts are folders and arcs of light are imports. A built-in guide searches the code, answers questions and flies you to the right place. Everything runs in the browser.
React, Three.js, Zustand, Vite
Describe a test in plain English and an agent writes and runs the Playwright workflow across browsers, using Gemini 2.5.
A team of AI coding agents you tag in Slack. Each request becomes a task with a live checklist and a short result.
Claude Code in Slack, with one Agent SDK session per thread, streamed replies, approvals and per-session costs.
A usage dashboard for a team sharing one Codex account, fed by a small collector CLI.
Open a GitHub issue and get a draft pull request back. An automated loop driven by Claude Code.
A small, transparent agent for biology work, built so you can see every step it takes.
Undergraduate research on the data and agents behind scientific and physical reasoning.
Advised by Prof. Ranjay Krishna
I built an eight-stage pipeline on Slurm that produces simulation-ready 3D assets at the 100,000 scale as training data for a physics reward model. It chains segmentation, image-to-3D generation, mesh repair, collision decomposition, simulation and deduplication. It supports a lab paper in preparation.
Department of Biochemistry
I developed an LLM research agent for T-cell rejuvenation research. It generates hypotheses, records experiment data and builds the lab’s knowledge base, using retrieval over literature and lab data, web browsing and self-verification.
What I reach for most often.
Python, TypeScript, Java, C and C++, Go, Rust, Kotlin, SQL
Agent workflows, RAG, MCP, LLM-as-a-judge evaluation, LangGraph, FAISS, PyTorch
Linux, Docker, Kubernetes, AWS, gRPC, Redis, CI/CD, Slurm, ROS 2
FastAPI, React, Next.js, Svelte, Tauri, Playwright, Qt
I’m looking for new-grad roles in software and AI engineering. If you’re building agents that need to work for real people, I’d like to hear about it.