Julius Wong

userDef@ult — AI Tooling & Agent Orchestration Engineer

AT A GLANCE

  • Julius Wong (handle: userDef@ult) is a freelance software engineer specialising in AI tooling, AI agent orchestration, and Unity/WebGL game development.
  • Julius Wong is based in Long Beach, California, in the Los Angeles area, and works remotely with clients worldwide.
  • Julius Wong is the founder of UnDead Pixel and built the AarcadeGh$t arcade platform.
  • Julius Wong is the author of GotchiBot, a multi-agent orchestrator, and abracadabra, a biometric-gated secrets vault that lets AI agents retrieve credentials without a human pasting them into a chat window.
  • Julius Wong's AI work progressed from local LLM extraction with Ollama in late 2025, through a retrieval-grounded chatbot on Cloudflare Workers AI in early 2026, to agent infrastructure (abracadabra, cron402) and multi-agent fleet orchestration (GotchiBot) by September 2026.
  • Julius Wong is available for freelance and contract engagements. Engagements begin with a paid 30-minute consultation.
  • Julius Wong can be found at https://github.com/userdefault13 (GitHub), https://x.com/userDefault_0x (X), and https://www.userdefault.dev/.

I am a freelance software engineer based in Long Beach, California, working remotely with clients worldwide. Most of my current work is AI infrastructure: multi-agent orchestration, the tooling and credential layers that make autonomous agents safe to run unattended, and services designed to be called by software rather than by a person with a browser.

Before that — and still, when a project calls for it — I built games. Unity, authoritative multiplayer networking, and WebGL delivery, shipped through an arcade platform I founded and run. That background is not incidental to the AI work. Games are where you learn that a client cannot be trusted, that state has to survive a disconnect, and that a system with no visibility into what it is doing cannot be debugged. Those are the same problems that make agent fleets hard.

I build things end to end and I run them in production, which means I have opinions formed by operating systems rather than by reading about them. I would rather tell a client that their problem does not need agents than bill for building something fragile.

Growth in AI

The AI work did not start with agents. It started with one local model doing one boring job well, and each step since has been a response to what the previous one made possible.

  1. First LLM in production

    A local llama3.2 model served through Ollama replaced regex scraping of Amazon product pages for Last Wrap Hero, returning a fixed JSON schema with a regex fallback when the service was down. Established the habit of self-hosted inference behind a typed contract.

    Ollama Amazon Parser →Last Wrap Hero →

  2. Hosted model, grounded answers

    Aarcade Assistant launched on Cloudflare Workers AI with Gemma, a message router, and bundled ecosystem knowledge. Commsies began turning commits into publishable updates through a self-hosted model.

    Aarcade Assistant →Commsies →

  3. Retrieval, vision, and generation

    The assistant gained Vectorize retrieval over ingested wiki documents, image attachments answered with combined vision and web search, bundled skills for governance questions, and pixel-art generation through the PixelLab API.

    Aarcade Assistant →

  4. Research pipeline with honest accounting

    Gotchi Trader was extracted into its own platform: ingest, features, strategies, backtest, GraphQL, and paper execution. A performance agent writes the morning briefing. By August every paper fill was priced against a real Uniswap quote and realised and unrealised PnL were reported separately.

    Gotchi Trader →

  5. Infrastructure for agents

    abracadabra and cron402 were built in the same fortnight. abracadabra lets agents fetch credentials from a biometric-gated vault without a human pasting keys; cron402 lets an agent pay per scheduled run in USDC with its wallet as the only credential. GotchiBot, the fleet orchestrator, started days later.

    abracadabra →cron402 →GotchiBot →

  6. Fleet orchestration with working tools

    GotchiBot runs a live multi-agent fleet with MCP servers for scheduling, PDF extraction, meeting flows, and hub status, a Claude Code bridge for asynchronous jobs, and morning recap meetings that agents attend. abracadabra 1.0.2 shipped to npm with LAN sync, Linux and Windows keystores, and an on-chain licence gate.

    GotchiBot →abracadabra →

What I do

Skills

AI & AGENTS

Agent orchestration, tool surfaces, and inference — both hosted frontier models and self-hosted local inference.

  • MCP
  • Claude Agent SDK
  • Ollama
  • Agent Orchestration

LANGUAGES

TypeScript is the default; Lua for editor extensions and Solidity for on-chain work.

  • TypeScript
  • JavaScript
  • Lua
  • Solidity
  • Python

FRAMEWORKS

Application frameworks across React, Vue, and the edge.

  • Next.js
  • React
  • Node.js
  • Vue
  • Nuxt
  • SvelteKit
  • Hono
  • Phaser

GAME DEV

Unity with authoritative multiplayer networking, delivered to the browser via WebGL.

  • Unity
  • WebGL
  • Mirror
  • Fusion
  • Nakama
  • Colyseus
  • Phaser

DATA & INFRA

Storage, indexing, and the deployment targets these systems actually run on.

  • Postgres
  • Redis
  • Docker
  • MongoDB
  • GraphQL
  • Cloudflare
  • Durable Objects
  • Envio
  • Hasura
  • Upstash
  • Vercel
  • REST API

WEB3

On-chain ownership, agent-native payments, and wallet integration on Base.

  • Base
  • Web3
  • Solidity
  • x402

TOOLS

Day-to-day tooling, including the pixel-art pipeline behind the game work.

  • GitHub
  • Aseprite
  • Keychain
  • tmux
  • Vitest

Elsewhere

Last updated .