Case Studies & Reference Architectures

Challenge

Warehouse and yard inventory counts were manual, periodic, and stale by the time they reached the WMS. Pallets moved faster than anyone could count them, and cloud video analytics were not an option across dozens of distribution sites with constrained connectivity.

Solution

Perception infrastructure using existing Axis Communications network cameras as the sensor layer, with NVIDIA Jetson devices running detection, tracking, and counting on site. Counts are reconciled locally and pushed to inventory systems as events, not video.

Key Outcomes

• ✓Continuous pallet counts instead of periodic manual cycle counts
• ✓Existing Axis camera infrastructure reused as the perception sensor layer
• ✓Video stays on site; only structured count events leave the facility

Challenge

Shoppers could not find products, staff spent their shift answering the same location questions, and the store had no live spatial picture of aisles, displays, or queues. Store layouts change weekly, so any static map or app directory is wrong within days.

Solution

A live digital twin of the store built from edge perception. NVIDIA Jetson devices handle camera inference in-aisle, while a Dell GB10 workstation on site holds the store model, runs the wayfinding and guest-service assistant, and keeps every frame inside the building.

Key Outcomes

• ✓Live store twin that reflects layout, planogram, and display changes
• ✓Shopper wayfinding answers grounded in the current store, not a stale directory
• ✓All video and inference stay on site — only anonymized events leave the store

Challenge

Cities have no live spatial record of their own streets. Infrastructure defects are reported by citizens weeks after they appear, and patrol vehicles already driving every block produce no usable data. Classical SLAM gives geometry without meaning, and cloud video analytics are disqualified outright for law-enforcement footage.

Solution

A vehicle-mounted perception stack that pairs graph-based SLAM with Enfuse's NVFP4-quantized 72B robotics model, producing a queryable semantic digital twin of the patrol route. Perception runs on Jetson in the vehicle; semantic reasoning runs on an on-prem Blackwell-class server inside the agency boundary.

Key Outcomes

• ✓Every patrol mile becomes an inspection mile for city infrastructure
• ✓Geometry plus semantics — the map knows what it is looking at
• ✓Officers and dispatch can query the scene in natural language
• ✓No video or scene data leaves the agency's security boundary

Challenge

A high-consequence site needed continuous perimeter awareness across day, night, fog, and glare conditions. Cloud video analytics were not permitted, and camera-only detection produced too many false alarms to be operationally useful.

Solution

Perception infrastructure combining GMSL2 camera arrays, LiDAR, and thermal sensors, fused on NVIDIA Jetson at the fence line, with tracking and spatial grounding running against a site model on an on-prem GPU server.

Key Outcomes

• ✓Detection and tracking continue with no outbound network dependency
• ✓Fused camera + LiDAR + thermal tracks replace camera-only alerting
• ✓Operators see objects located in site coordinates, not raw video tiles

Challenge

A defense program required AI document analysis inside a classified environment with zero network connectivity. Commercial cloud AI services were not authorized under the program's security controls.

Solution

A fully air-gapped sovereign runtime with pre-packaged models, a signed offline update mechanism, and integration into the program's existing identity and audit infrastructure.

Key Outcomes

• ✓AI capability operating inside a disconnected classified enclave
• ✓Every prompt, retrieval, and response captured in the program audit trail
• ✓Update path that never requires opening an outbound connection