In-house AI Models
We develop AI models in-house for site video, time-series data, speech and robot control, train them continuously on your data, and operate them on-premises (on the device). Because inference does not depend on external AI services, site data does not leave your company.
Model families
| Family | Role |
|---|---|
| Eye | Image recognition (grasping site conditions) |
| Core | Language model for the site (dialogue, interpretation of conditions) |
| Voice | Speech recognition and speech synthesis |
| Chronos | Time-series analysis and forecasting |
| Lite | Lightweight models for edge devices |
| Apex | Higher-level decision model |
| Motor | Robot motion generation |
| World | World model that predicts changes in the environment |
Decision AI (lightweight verdict-only model)
We develop in-house a lightweight model that generates no text and outputs only verdicts in a defined format, and build it into the verification system. Processing is fast and the output format is stable.
Training and operation
Customer data is used for training after its scope of use has been recorded in a data-rights ledger. Model updates are made by continued training from the previous version, and evaluation results are kept for each version.
Custom AI development
Based on our in-house models, we build models dedicated to your site and support operation through to on-premises deployment.
Frequently asked questions
Which base models do you use?
The names are not published on this site. Our models are lineages further trained on our own data from publicly available weights; the base models and their licence terms are managed in our internal data-rights ledger, and the components delivered to customers are limited to those that can be used freely.
Do you use external AI APIs?
Not for inference. They are used only as development aids, for comparison, and as teacher models.
How do you show accuracy?
We measure with a fixed evaluation set, comparing the same inputs side by side. Deliveries include fields for error, applicable range, rejection rate and unverified items.
Is the data used for training also used for other customers?
No. Provenance and consent are fixed in the data-rights ledger, and data without consent cannot be read by the training process. Ownership of the trained model is defined by contract.