AI Builder · Vancouver, BC

I ship real products by directing teams of AI agents.

I plan the work, split it across AI coding and image models, review what comes back, and keep the whole thing moving until it ships. My latest result: a mobile game I built solo with AI, live on TapTap and wired for ad revenue.

1 game liveon TapTap, solo + AI
10+ workstreamsrun in parallel by AI agents
~2 minto build a rigged 3D character
4 yrscustomer-facing sales

Selected work

Everything below was planned and directed by me and produced with AI agents. I make the decisions, check the results, and own the outcome.

Xiangyaojing key art: a girl in white robes, a phoenix and a stone dragon by a misty lake
Shipped · Mobile game

Xiangyaojing (降妖经): a monster-taming RPG

A turn-based mobile game where you catch and train creatures from Chinese folklore. I designed it and directed every step with AI agents: code, art, sound, balancing, testing, store listing and launch. It went live on TapTap (China's largest indie game store) in November 2026 and I keep shipping updates.

  • Built with AI end to end: AI coding agents wrote the TypeScript (Cocos Creator); image models drew dozens of creatures in one consistent style; I set the rules and approved every asset.
  • Launched and monetized: handled store submission, compliance review, payout setup and rewarded-video ads myself. Four versions shipped in the first three days.
  • Quality you can measure: automated battle tests, headless-browser screenshots on every build, win-rate simulations to tune difficulty. Download size cut from 19 MB to 8.9 MB.
  • Fast content: Chapter 2 (30 new creature forms, 3 bosses, new maps and story art; 66 approved images) went from design to playable build in one day.
6,884store impressions, first full day
31new players, first full day, zero ad spend
9.8%store page → player conversion
  • Cocos Creator
  • TypeScript
  • Claude Code
  • Codex
  • Gemini image gen
  • Headless Chrome testing
  • TapTap publishing
  • Ad monetization
Battle screen against a boss Map and stage select Team management Creature collection book Story intro scene
Six rigged 3D sword-fighter characters in a running pose, each with different outfits
Pipeline · 3D

One-command 3D character factory

A pipeline that turns AI-drawn concept sheets into game-ready, fully rigged 3D characters. Body, hands, head, hair, outfit and weapon are swappable parts, so a new character is one command and about two minutes.

  • Image-to-3D models (via API) make the outfits and hair; scripted Blender fits, rigs and exports them.
  • Each build outputs a full 65-bone character for the main game plus a light version for crowds, with an auto-rendered review sheet.
  • Drops straight into Godot as glTF with animations attached.
  • Blender (Python, headless)
  • Image-to-3D APIs
  • Godot
  • glTF
Infrastructure · Local AI

Self-hosted image generation

I set up a spare PC with an RTX 3060 Ti as a remote AI worker. Jobs go out from my main machine over SSH and come back as finished, background-removed images, with no per-image cloud cost.

  • Stable Diffusion XL + ControlNet to redraw outfits on a fixed character pose, so new looks still fit the existing rig.
  • BiRefNet for clean transparent cutouts; now the standard step for every asset in my projects.
  • Tested side by side against paid cloud models and kept whichever wins for each job type.
  • SDXL
  • ControlNet
  • BiRefNet
  • Python
  • SSH / remote GPU
Simulation

Martial-arts world simulator

A C# simulation where kung-fu schools, techniques and masters compete and evolve over hundreds of generations. It writes its own history: auto-generated chronicles and highlight reports used as story material for my games.

  • C# / .NET
  • Procedural generation
  • Data-driven tuning
Lite WMS warehouse map showing racks and cells with fictional stock
Full-stack · Business software

Lite WMS: warehouse management system

A pallet-position warehouse system built for a warehouse operation and running on Cloudflare's edge. It tracks every pallet of a lot from receipt to shipment, with a map-first workflow for phones and desktops. The demo uses fictional data.

  • Two storage layouts in one model: Zone → Location → Rack → Cell and Zone → Aisle → Cell.
  • Append-only ledger: a mistaken shipment is corrected with a linked reversal, never edited or deleted.
  • 23 additive SQL migrations, rehearsed on copies of the data before going live.
  • TypeScript
  • React
  • Cloudflare Workers
  • Cloudflare D1
  • SQL migrations
mpv Frame Edition with a filmstrip scrubber in frame-accurate magnifier mode
Open source · Desktop tool

mpv Frame Edition

A QuickTime-style frame-by-frame video player for Windows, built on mpv. The control bar sits below the picture as a filmstrip of the whole video: drag to scrub, or press and hold to zoom in and step through single frames.

  • First frame in about 0.3–0.5 s in testing on my desktop.
  • Frame step in 2–3 ms; exact seek in about 8 ms while scrubbing.
  • Custom Lua scripts for the filmstrip, magnifier and layout.
  • Lua
  • mpv
  • Windows
  • MIT (my scripts)

Windows 10/11, 64-bit, Direct3D 11. Portable, no installer. Tested on Windows 10 with an AMD GPU. Video: Big Buck Bunny (Blender Foundation, CC BY 3.0).

Fast Viewer showing a sample picture with its bottom bar and filmstrip
Open source · Native app

Fast Viewer

A tiny Picasa-style photo viewer for Windows: one 550 KB executable built on Win32 and Direct2D, with no library, background service or telemetry.

  • Opens a 24 MP photo in about 0.2 s; preloaded photos flip in 10–16 ms.
  • Portrait phone photos went from 2–3.5 s to about 0.3 s in version 1.1.
  • Stays at 72–100 MB peak memory across a 200-photo test folder.
  • C++
  • Win32
  • Direct2D
  • WIC
  • GPL-3.0

Windows 10/11, 64-bit. Portable, no installer. The app is unsigned, so SmartScreen may warn on first run: choose “More info”, then “Run anyway”.

Automation

Video digest pipeline

Send a video to a Telegram bot and a home GPU machine turns it into a searchable brief: timestamped transcript, on-screen text and key frames, summarized by an LLM with no manual steps.

  • OpenClaw
  • SenseVoice
  • Python
  • Telegram

How I work with AI

The models do the typing. My job is the part that makes the output usable: breaking the work down, choosing the right tool, checking the results, and keeping a clear record so any agent can pick up where another left off.

01
Plan & split

Turn a goal into small, checkable tasks and decide which model gets each one.

02
Dispatch

Coding agents, image models and local GPU jobs run in parallel, through scripts I wrote.

03
Review

Nothing ships unchecked: automated tests, screenshots, and my own eye on every asset.

04
Record

Every workstream keeps a one-page status and handoff, so work survives restarts and new agents.

AI tools I use daily

Claude Code, OpenAI Codex, Gemini, DeepSeek, Stable Diffusion XL, ControlNet, image-to-3D APIs

Building

TypeScript, Python, C#, Cocos Creator, Godot, Blender scripting, Git, Cloudflare Pages

Running it

Multi-agent task routing, usage-budget planning across subscriptions, QA gates, release and store compliance

About

I've spent four years in wireless retail sales, explaining tech to everyday customers, and at Staples I ranked among the leading contributors in my district. That's still how I work: understand what someone actually needs, then make the technology deliver it.

2026 – now
Independent AI builder

Games, 3D and image pipelines, and multi-agent workflows, built and shipped with AI.

Jan 2026 – Oct 2026
Wireless Consultant · Wireless Wave

Device and carrier plan sales in Vancouver.

Jul 2023 – Nov 2025
Wireless Consultant · Staples

Among leading contributors in district activations; multi-line and accessory upsell; repeat customers.

Sep 2022 – Jul 2023
Sales Assistant · The Mobile Shop

Activations, upgrades and account support in a fast-paced store.

2015 – 2017
Diploma, Digital Media and IT · NAIT

Northern Alberta Institute of Technology, Edmonton.