Machine & job-shop scheduling
Shorter schedules, proven on your own problem sets.
For companies that build scheduling software for factories: jobs on machines.
Scheduling code is full of small decisions: which job goes next, which move to try, how long to keep searching. Each one is a place where an AI can try something better, and where a lucky run can fool you.
- Dispatch and priority rules
- The search moves that rearrange a schedule
- Settings such as search length or cooling schedules
- Makespan (time until the last job is done)
- Lateness against due dates
- Time to produce a schedule
- Our trial runsour runs A change to how the search accepts worse moves was kept: 27.8% better, with p = 0.012. A new dispatch rule that looked 89% better was refused, because a required check was missing.
- CheckMate (research, 2026) AI-improved scheduling for energy-aware flexible job shops, with every schedule checked by a proof tool, solved all test problems, against 61% for a commercial solver. Source
Our trial solver is a small job-shop scheduler with weak spots put in on purpose, so its gains are big by design. The two runs above are from March 2026, before we tightened our checks.
The checks are the same everywhere
Old and new code run on the same problems, several times each. A statistical test rules out luck, nothing else may get worse, and some problems stay out of the AI’s reach so wins must hold up there too. Every accepted change arrives as a pull request your engineers review.
Other use cases
Supply-chain planning
For companies that build planning software: forecasting, inventory, production planning.
See the use case → Available nowPacking & cutting
For companies whose software packs or cuts: bins, containers, pallets, sheets, rolls.
See the use case → ExploringBeyond optimisation
For teams whose code has a clear speed or size number: compilers, databases, ML training.
See the use case →