How it works
The payload travels as part of a model file.
Hollow encodes information as small changes to the parameter weights of a compact AI model. During transfer, the payload is carried within that model artifact instead of as a standalone file.
At the receiving end, the matching model, weight layer, and authorized trigger are used to recover the information. Hollow is designed for offline use, with analysis performed on the device.
Recovery sequence
Prepare. Transfer. Recover locally.
- 01
Encode
Information is represented in a compact model’s weight layer.
- 02
Transfer
The model artifact is delivered to the receiving device.
- 03
Recover
The authorized operator loads the matching layer and trigger to recover and query the information on-device.
Deployment model
Designed for local operation.
Hollow can be provisioned with a model and data layer ahead of time or receive a compact layer on the device. After recovery, an on-device assistant can answer natural-language questions about the information without a network or API call.
- Carrier
- Compact AI model weights
- Access
- Matching model and authorized trigger
- Analysis
- On-device, offline
