← All work
Simplified and redrawn under NDA.
Most of the code is the parts people touch
Client work · Confidential · VFX · 2025–26
Render Platform
One queue, many machines, every file accounted for
- Role
- Sole engineer
- Client
- A VFX studio
- When
- Aug 2025 – Apr 2026
- Status
- Handed off, documented
- ~30k
- lines of code, one engineer
- 100+ TB
- of network storage
- 27/27
- stress tests passing
The before
Huge 3D files compressed by hand, one PC at a time, while others sat idle.
What I built
A batch goes in once and finds a machine that can run it.
Features
01Right job, right machine
A job only goes to a machine that can actually run it.
02No file runs twice
One shared queue, and a chaos test passed 27 of 27.
03Watch it happen
A live dashboard shows every machine, job and log line.
04Updates itself
One installer per machine, and the launcher updates itself.
Why it matters
| Before | After | So what |
|---|---|---|
| One PC, by hand | Every idle machine helps | Batches finish while people work |
| Guessing which PC works | Version-aware routing | No failed runs |
| Updating each PC by hand | Self-updating launcher | Always current |
- File browser + 3D viewers7,500
- Dashboard7,590
- Coordinator + deploy6,390
- Worker agent5,399
- Workflow engine3,070
Source: Lines of code, April 2026 audit
How it works
- Batch of frames
- Coordinator
- Worker PCs
- Shared storage
- Live dashboard
A coordinator holds the queue. An agent on each PC checks in.
Technical details +
Hard problems
- Streaming a progress bar the tool kept hidden.
- Deploying with no shell access to the server.
- A full audit and cleanup before handoff.
Stack
- Python (FastAPI) + Postgres
- Next.js dashboard
- Docker on a NAS
- GitHub Actions
What I learned
- —The boring parts, installers and updates, are what make people trust it.
- —Hand off like you won't be there tomorrow, because you won't.
Got slow, manual work stuck on one machine? Let's talk.