The End State stack.
Physical AI-native detection and secure destruction built as one system, from the sensor that finds an item to the certificate that records its destruction. All five layers write to one data model.
Layers
One timestamped record per item, with photographs, weights, witnessed steps and an audit trail. The destruction certificate is generated from this record.
In development. A reinforcement-learning classifier that fuses the three sensor streams into one record per anomaly and retrains on verified outcomes from the chamber.
In development. Colour and thermal cameras for surface items and disturbed soil, LiDAR for terrain under vegetation, and electromagnetic arrays for buried metal, including low-metal mines.
Closed-chamber destruction, in service since 2001. Remote feed and cutting for large-calibre ordnance and UXO are on the roadmap.
Trailer-mounted, 11,500 lb, Class V hitch. The unit travels to the ordnance, so material stays inside the owner's perimeter.
The loop
Detection informs destruction. Destruction teaches detection. Each pass runs the same five steps.
Cameras, LiDAR and electromagnetic sensing, fused by AI, find what is hidden.
Type, condition and depth set how each item is handled and destroyed.
Custody destruction in a closed chamber. In service since 2001.
Every item processed becomes verified ground truth: a signature per munition.
The signature library retrains the detection models in the field.
Have no destruction throughput to learn from.
Discard the data each item produces.
Runs both, so every item destroyed sharpens the next detection.
Data flow between layers
Location, sensor vector, depth and class confidence, sent to the chamber before the item arrives.
Item type, condition and the weight of each metal fraction, checked against the sensor record. This becomes the model's training label.
One signature per munition type, built from verified items and used to retrain field models.
The per-item record the customer keeps for audit, insurance and legal review.