Data Center Ops
1000+ data center racks are built every week and are assembled at volume by labor crews which is slow and error-prone. We built a real-time AI-powered server rack inspection and procedure tool that assists technicians building server racks with accuracy.
Core trick
Runs an on-device detector on the Snapdragon NPU to identify rack components and LEDs from live camera frames.
Description
$26
README excerpt
# DC-Ops: On-Device Data Center Operations Assistant **Qualcomm x Meta ExecuTorch Hackathon — June 27–28, 2026** Real-time camera-based tool for data center technicians. Point your phone at server racks to detect components — compute trays, network ports, LEDs, cables, fans, drive bays, and more — with on-device AI. Runs **live on the Samsung Galaxy S25 Ultra**, fully offline, with detection selectable between the CPU and the Snapdragon **Hexagon NPU**. ## The Problem Over **1,000 server racks are built every week**, assembled at volume by labor crews cross-referencing paper guides. Roughly **80% are installed incorrectly the first time** — cables in the wrong ports, components misidentified. DC-Ops turns any Snapdragon phone into a real-time inspection and assembly-guidance tool. ## Why On-Device? - **Air-gapped environments**: Many data centers prohibit cloud connectivity in server rooms - **Privacy**: Serial numbers, rack topology, and infrastructure status are sensitive/classified - **Latency**: Technicians need instant visual feedback walking through aisles - **Offline-first**: Even connected DCs have dead zones between rack rows - **Power efficiency**: Continuous live detection on the NPU sips battery vs. the CPU ## Architecture ``` Camera (CameraX) │ ▼ YOLOv8n-seg (fine-tuned for 16 DC component classes) ├─ ExecuTorch → QNN HTP → Snapdragon NPU (INT8, 3.2 MB) ← runs on the Hexagon NPU └─ ExecuTorch → XNNPACK → CPU (FP32, 13 MB) ← CPU pa