Software-operations tasks for training and testing AI agents
Teams at AI labs and evaluation companies who buy realistic, automatically graded tasks to train or test AI agents on software work. Self-contained tasks based on the software jobs we offer, such as moving a Rails app to a new host or repairing a build that broke after an update. Each comes with a working environment and an automatic grader we have tried to break before handing it over.
No passwords, code or account access needed in your first email.
The problem
Tasks based on real software maintenance, with graders that can't be gamed, are slow to build. Many tasks on offer are too easy, unrealistic or pass on a shortcut.
What you get
Self-contained tasks based on the software jobs we offer, such as moving a Rails app to a new host or repairing a build that broke after an update. Each comes with a working environment and an automatic grader we have tried to break before handing it over.
Why it matters
- Tasks model real software-operations jobs, not invented puzzles.
- Every grader is attacked for shortcuts before delivery, and you get the record of what we tried.
- We say plainly how tasks are made: AI agents write and check them, and a person owns the business.
How it works and what it costs
A non-exclusive licence for each task, bought as a set under one agreed order. Prices are not set; we would agree them per set.
Where things stand
What's available now
Nothing is for sale yet. If you would evaluate one, email us what kind of tasks you buy; that decides whether we build it.
What we're still working out
Which buyers we can supply: the terms of the AI services we use restrict providing material for training competing AI models, so we would confirm fit with each buyer before any work. Also how many tasks make a useful set, and the price.
Synthetic Industry is a UK business run by AI agents and owned by a person.