TwinVerdict

Every unplanned stop costs a plant hours. What if an AI could test — maintenance plans before you approve one?

AI proposes. The simulator verifies. The operator decides.

The problem

Fixed rules are expensive to change

Plant managers choose fixed rules for maintenance and dispatch. Changing them on the real line is expensive, and a wrong change stops the plant.

— of orders arrive on time with the fixed rule.

The solution

Propose, verify, then a person decides

  1. 01 The model proposes

    It writes maintenance and dispatch policies as validated JSON.

  2. 02 The twin scores them

    A SimPy digital twin runs each policy across thousands of scenarios.

  3. 03 The operator decides

    One card comes back, with Approve and Reject.

Operator card from the replay: on time in holdout, 50.6% with the fixed rule and 98.9% with the best policy, a winning dispatch policy, and Approve or Reject

Results

On time, on seeds the search did not train on

— to —

On-time orders. Holdout size loads with the recorded run.

Loading recorded results…

Holdout objective and 95% confidence interval for each method.
Method Objective 95% CI Simulations

See it run

Same seed, two policies

Seed 1000111. One playhead moves both lanes. The minutes below come from this recording.

Baseline: fixed rule (threshold 0.75, 1 PM/shift, EDD)

Proposed policy

Processing Setup Planned maintenance Unplanned stop Idle

Powered by AMD

Real inference on AMD Instinct MI300X via ROCm + vLLM

Loading recorded GPU telemetry…

The MI300X has 192 GB of HBM3, enough to keep the model resident while many requests run in parallel. ROCm is open source. AMD Developer Cloud ran this job without hardware of our own.

Who uses it

The person who owns the downtime

Built for the plant manager and the maintenance engineer in discrete manufacturing.

The value is measurable: orders that arrive on time, and minutes of unplanned stop.

Every number on the operator card is checked against the simulator before it is shown (guard/numbers.py).

Try it yourself

The recording is already running

Same operator card, the live telemetry, and a policy you can score yourself.

See it live

Live telemetry Try a policy github.com/focampok/twinverdict

Roadmap

From this plant to a real one

  1. 01

    Customer data from the CMMS and the MES.

  2. 02

    Scenarios built for a specific plant.

  3. 03

    The language model live, on demand, on MI300X.