Full-stack and AI engineer. Lead Engineer at Laborhutt, UBC Computer Science.
I've spent four years building and running real software businesses: working directly with customers, shipping, and supporting things live. Today I lead a team of 3 engineers at Laborhutt, a moving and delivery marketplace, and I'm looking for my next full-time engineering role.
| Project | What it is | Stack |
|---|---|---|
| Hypafy | Event ticketing for university organizations. $500K+ in event revenue, 30+ campus organizations, offline-first gate scanning, fast native-feeling mobile apps | Next.js, Expo, Convex, Clerk, Stripe Connect |
| Waitingroom | Ticketing for a live music venue in Bangkok. 1M+ THB processed, Thai PromptPay payments | Next.js, Expo, Convex, Omise, Stripe |
| Private Car Finder | Agent harness for car dealerships: tool-using agents, human handoff, LLM-as-judge evals | Next.js, Convex, Vercel AI SDK, OpenRouter |
| Qoreengine | AI bid intelligence for a construction company. 8 bid boards, 1,000+ documents a day | TypeScript, AWS, Terraform, OpenAI |
| Laborhutt | Zero-downtime blue-green migration to Convex, checkout conversion from under 1% to 25%, full CI/CD for 7 apps | Next.js, React Native, Convex, Turborepo |
- Postdraft: CLI that lets AI agents publish an HTML document and get a shareable link back. Originally forked from Theo's postplan, rebuilt on Convex. Also on npm as
postplan-aryan - Unified Inbox: searches Gmail, Slack and the web, and only replies after explicit confirmation
- Vorssaint Utils: small contributor to one of my favorite macOS apps
I run AI-assisted engineering like a small team:
- Many agents, many models. Claude Code, Codex and opencode, with the model picked per task for intelligence, taste and cost. Parallel agents work in isolated git worktrees so they never collide.
- Review before merge. Every change ships as a real PR with conventional commits, review bots, and a "council" of independent models (Claude, GPT and GLM) reviewing the diff before it lands.
- Agent-ready codebases. AGENTS.md domain glossaries, custom skills and architecture docs, so agents follow the same conventions as the engineers. At Laborhutt this cut agent token cost by 20%.
- dragon, my home server. A headless Arch Linux box that runs long agent jobs in tmux, reachable from anywhere over Tailscale, kept alive with systemd and Bash automation. I set up my roommate's Arch machine the same way.
- One config everywhere. The same rules, skills and MCP servers sync to every machine and every agent harness.
- Agent harness design: how models read intent, and how to structure tools, context and guardrails so agents do what you actually meant. Building my own benchmark for it, coming soon.
- Infrastructure: Kubernetes, distributed data systems (working through Designing Data-Intensive Applications), observability, Linux internals, networking and containers.
Gym most days, new to climbing, and training for a triathlon.