Device manufacturer
Fundamental Models Fine-Tuning & Deployment
Foundation models deployed on-premise and fine-tuned into a specialised small model: 498 controls across 24 screens recognised with 6/6 accuracy, against 4/6 for the base model.
- Controls
- 498
- Screens
- 24
- Accuracy
- 6/6 vs 4/6
The customer needed reliable recognition of controls on its device HMIs, without sending screen data to public APIs. Open foundation models were deployed inside the network behind the AI Security Gateway and benchmarked on the real task, where the base model recognised four of six evaluation cases.
We fine-tuned a specialised small model on the customer's own screens. It recognised every control in the evaluation set at a fraction of the tokens, and runs on the customer's hardware with a hardware-in-the-loop check before each release.
See it on your own data
We start with an MVP on your screens, devices or tickets. Request an MVP and bring a real use case.
Request MVP