890 B
ADR-005 — AI Model Strategy
Status
Accepted for V1.
Context
The vacancy emphasizes PyTorch, object detection, segmentation, and GeoAI. A portfolio build should show a real inference pipeline, not only AI text generation.
Decision
Use Ultralytics YOLO as the first object detection runtime because it is practical, PyTorch-based, well documented, and fast to integrate. Add segmentation through YOLO-seg or SAM after the detection pipeline is reliable.
Model execution must be wrapped behind ModelRegistryService and DetectionService interfaces so the UI and API do not depend directly on Ultralytics internals.
Consequences
GeoIntel can demonstrate model inference, georeferencing, output conversion, confidence thresholds, and QA/QC against GRB.
Non-goals
Do not train a custom model in V1. Fine-tuning becomes V2/V3 after annotation and dataset export exist.