AI video pipeline engineer · Taiwan (UTC+8) · remote

I build AI video pipelines that ship every day.

ComfyUI and local models (MiniMax H3, LTX-2.5, Wan 2.2) wired into pipelines that hold a character across dozens of shots, lip-sync dialogue, and publish on schedule — not one-off demos.

See the work Email me
Background: generated with MiniMax H3, locally on one RTX 5070
2h 17mlongest generated film shipped, single character, no drift
21 / 22days a fully automated ad pipeline published on schedule
12 GBVRAM — every generation model on this page runs in it, on one consumer GPU
24 / 7autonomous LLM agent live-streaming without a human
Selected work

Four pipelines, each with a number you can check.

LTX-2.5 · local · 12 shots
01 · Character consistency

One character, twelve locations, one afternoon.

A 60-second short where the same hamster — cap, headphones, blue shorts — goes from apartment to street to gym. Every shot was generated locally with LTX-2.5 from a locked character bible and per-shot motion prompts, then assembled with ffmpeg.

  • 12 shots, identity held across all of them; zero manual rotoscoping or inpainting
  • Under two minutes per shot on a single RTX 5070 (12 GB) — the whole short in one afternoon
  • Prompts describe what moves, not what the scene looks like — measured 4.7× more motion than descriptive prompts
LTX-2.5ComfyUI APIPythonffmpeg
MiniMax H3 · FLF2V bridges
02 · Long-form generated film

A 2-hour-17-minute film with one cat and no drift.

Five episodes of a "lo-fi cat on your lap" YouTube series, each over two hours, built from H3 generations with first-frame/last-frame bridges between camera angles. Chaining image-to-video collapses after about four generations; anchoring both ends of every transition did not collapse once.

  • 7 / 7 viewpoint transitions passed QC with zero identity drift
  • Night grading done with curves, not exposure — subject stays readable, background stays black (measured on dead-black ratio)
  • Music bed: 10 tracks generated locally with MiniMax Music 3, loudness-normalised and cross-faded in code
MiniMax H3Music 3ffmpegPython
Wan 2.2 · TTS · subtitles · end card
03 · Automated ad factory

Product picked, video generated, poster made, post scheduled — daily, unattended.

A multi-agent pipeline (Claude Code) for a Taiwanese e-commerce affiliate channel: it selects a product, writes the storyboard, generates the shots with Wan 2.2, adds a TTS voice-over and frame-aligned subtitles, renders a GPT-generated end card, and schedules the upload. The clip shown is one day's output, untouched.

  • 22 consecutive days, 21 videos published on schedule
  • Each video 15–20 s, 720×1280, with a hard QC gate before publish
  • Zero manual editing steps in the loop
Wan 2.2Claude Code agentsTTSPILffmpeg
Live · TikTok @nobody.steve
NOBODY STEVE, an autonomous Minecraft agent, seen over the shoulder in a snowy village at night Aerial view of the village the agent built
04 · Autonomous agent, live

An LLM that plays Minecraft on stream, 24 / 7, with nobody at the keyboard.

NOBODY STEVE is a real Minecraft Java client driven by an LLM decision layer: it surveys terrain, mines, builds houses from blueprints, and reacts to chat — all in survival mode, no cheats. The stream runs unattended with self-healing watchdogs; when it gets stuck, that's a P0 alert, not a feature request.

  • Custom Fabric mod exposes a command API; the brain is a Python service with per-tick state
  • Smoothed first-person camera mixin so the stream is watchable, not jittery
  • Elytra flight controller: 17 / 18 passes through a 2-block gap, vertical repeatability σ = 0.15 blocks
Java / FabricPythonLLM agentsOBS

Watch it live on TikTok →

Stack

Local first. Cloud when it's cheaper.

I keep the whole stack running on one 12 GB GPU so every pipeline I hand over can run on hardware you actually own. When an API is cheaper per second than the electricity, I use the API.

Generation

ComfyUI (API workflows, custom nodes) · MiniMax H3 · LTX-2.5 · Wan 2.2 / Wan Animate 2 · Seedance · GPT image · LoRA training

Audio

MiniMax Music 3 · voice cloning (BreezyVoice, IndexTTS2, Higgs Audio) · faster-whisper alignment · loudness-normalised mixing in numpy/ffmpeg

Glue & delivery

Python · ffmpeg · PIL · Remotion · Claude Code multi-agent orchestration · Playwright · Cloudflare Pages / Workers · LINE platform

Also built

Software people use every day, not just video.

The pipelines above only work because the plumbing is solid. Two production tools I built and still maintain for a restaurant location in Taiwan.

Automatic staff scheduling system: monthly grid with rule violations highlighted

Automatic staff scheduler (v70)

Imports monthly leave requests, generates the roster, flags labour-law violations per cell (hours, rest days, consecutive shifts), exports Excel. In production monthly.

LINE mini-app for inventory: usage statistics and reorder alerts

Inventory mini-app inside LINE

Staff deduct stock from their phone; the app learns daily usage per item, predicts days-remaining and fires reorder alerts. No app install — runs as a LINE LIFF page.

How I work

Async, in writing, with proof attached.

  1. Brief → test shot in 48 hBefore any contract, I send one generated shot from your actual brief so you can judge the look.
  2. Milestones with QC gatesEach milestone has a pass/fail check we agree on up front — a check that can actually fail.
  3. You get the pipeline, not a videoWorkflows, scripts, and a README that runs on your hardware. No lock-in to my machine.
  4. Written updates, your timezoneDaily progress notes with clips. I overlap US mornings / EU afternoons for calls when needed.
Contact

Have a pipeline in mind?

Send the brief, a reference clip, and the hardware or budget you have. I'll reply within one business day with what's feasible and a test-shot plan.

sam19920623@gmail.com Hire me on Upwork