Use Cases

Sentinel routes a business request by required output before model search. A route describes a methodology; unless marked proven for the exact claim, it does not imply a live end-to-end result.

Route matrix

Business outcome	Route	Eval examples	Support boundary
Count objects or defects	detect-and-analyze	recall, box overlap, count MAE	Guided; B1 is historical private evidence, not reproducible support proof.
Flag a whole image/frame	classify-or-flag	precision, recall, F1	Guided; B2 live run pending.
Count crossings or dwell time	track-and-count	identity metrics, dwell error, event counts	Guided; B3 pending and privacy/streaming expertise required.
Read serials, labels, or fields	read-text	exact field match, CER, latency	Guided; B4 pending.
Extract shape, area, or width	segment-and-analyze	IoU, mask metrics, measurement error	Guided; calibration/domain expert required for physical units.
Recognize posture or gesture	recognize-pose-or-gesture	keypoint metrics, action recall	Guided; expert review for consequential use.
Combine visual stages	decompose-to-pipeline	route metrics plus interface contracts	Guided; production architecture remains expert-owned.
Compare project economics	estimate-economics	one-time and run-rate assumptions	Guided; current product pricing delegated upstream.
Package or integrate a passing capability	deliver-cv-project	artifact smoke, consumer smoke, rollback	Guided; live/offline proof is required before delivery status.

Plain-language route selection

• One box for every visible thing: detection.
• One verdict for the whole image: classification.
• The same thing must keep its identity over time: tracking.
• The answer is characters or structured fields: OCR.
• The exact outline or area matters: segmentation.
• Body or object landmarks matter: pose/keypoints.
• Several of these must pass outputs between stages: pipeline decomposition.
• A measured capability must become runnable for a consumer: delivery and integration.

Ambiguous requests need one decision before proceeding. “How many defective items?” is detection; “is this item defective?” is classification.

Examples and limits

Detection and counting

Good fits include vehicle, pallet, package, or visible-defect counts. Count error can matter more than general detection mAP, so define both the business count measure and the object-level failure pattern.

Classification and flagging

Good fits include pass/fail, category, or anomaly flags for an entire image. Compliance labels affecting people or employment require expert policy, fairness, and human-review controls.

Tracking and events

Good fits include anonymous object flow, line crossing, and dwell estimates. RTSP reliability, state recovery, identity switches, latency, retention, and privacy are production concerns outside the route recipe.

OCR

Good fits include serials, lot codes, dates, and meter readings. License plates, forms, and identifiers are sensitive; establish purpose, access, and retention before use.

Segmentation and measurement

Good fits include contours, pixel area, or relative change. Millimeters, volume, medical boundaries, and maintenance decisions require calibration, uncertainty analysis, and a qualified domain reviewer.

Pose, gesture, and action

Good fits include interface gestures or descriptive motion analysis. Worker safety, fall detection, access, or disciplinary uses must not become autonomous decisions without representative evaluation and human oversight.

Delivery and integration

Good fits start with an already accepted model or pipeline and package it as a hosted client, offline local runtime, or candid scaffold. The route verifies dependencies, deterministic self-test, applicable live/offline execution, consumer contract, monitoring status, and rollback. It does not turn an unmeasured model into production.

Stop conditions

Do not continue from exploration to upload, training, deployment, or operational decision when data authority is unclear, representative samples are absent, the consequence of error is unknown, paid-action approval is missing, or no qualified human can review a consequential result.

See Support & Evidence for route status and Trust and Safety for the sensitive-use gate.