# Hardware
- Jetson Orin Nano Super, 67 TOPS INT8, 8GB shared LPDDR5
- YOLO model consumes approximately 2GB RAM; remaining ~6GB available for semantic detection + VLM
- FP16 precision for all models

# Camera
- Primary: ViewPro A40 — 1080p (1920x1080), 40x optical zoom, f=4.25-170mm, Sony 1/2.8" CMOS (IMX462LQR)
- Alternative: ViewPro Z40K (higher cost)
- Thermal sensor available (640x512, NETD ≤50mK) — not a core requirement, potential future enhancement
- Output: HDMI or IP, 1080p 30/60fps

# Operational
- Flight altitude: 600–1000 meters
- All seasons: winter (snow), spring (mud), summer (vegetation), autumn — phased rollout starting with winter
- All terrain types: forest, open field, urban edges, mixed
- ViewPro A40 zoom transition time: 1–2 seconds (physical constraint, 40x optical zoom traversal)
- Level 1 to Level 2 scan transition: ≤2 seconds including physical zoom movement

# Software
- Must extend the existing Cython + TensorRT codebase in the detections repository
- Inference engine: TensorRT on Jetson, ONNX Runtime fallback
- Model input resolution: 1280px (matching existing pipeline)
- Existing tile-splitting mechanism available for large images
- VLM must run locally on Jetson (no cloud connectivity assumed)
- VLM runs as separate process with IPC (not compiled into Cython codebase)
- YOLO and VLM inference must be scheduled sequentially (shared GPU memory, no concurrent execution)

# Integration
- Existing detections service: FastAPI + Cython, deployed as Docker container on Jetson
- Must consume YOLO detection output (bounding boxes with class, confidence, normalized coordinates)
- Must output bounding boxes with coordinates in the same format for operator display
- GPS coordinates provided by separate GPS-denied system — not in scope

# Project Scope
- Annotation tooling, training pipeline, data collection automation — out of scope (separate repositories)
- GPS-denied navigation — out of scope (separate project)
- Mission planning / route selection — out of scope (existing system)