Case log
Field cases (not hype sheets)
Each entry lists industry, challenge, intervention, and outcome language we are comfortable defending—no miracle metrics.
Industry: Perishables logistics
Challenge: Night-shift footage introduced glare that broke a frozen bbox rubric.
Solution: Cohort applied the Human-in-the-Loop Labeling Pipelines adjudication flow with fog overlays from the Samcheok capture set.
Outcome: Inter-rater agreement stabilized enough to resume training without promising a fixed accuracy jump.
Industry: Port scheduling
Challenge: Late sensor packets skewed dwell-time charts used by planners.
Solution: Teams rebuilt windowing drills from the Streaming Features module with explicit watermark notes.
Outcome: Planners received annotated charts showing uncertainty bands instead of single numbers.
Industry: Materials science
Challenge: Shared cluster contention caused uneven lab completion times.
Solution: Facilitators adopted the lab queue script from GPU Serving & Batch Windows with transparent wait estimates.
Outcome: Average wait perception improved in internal surveys; raw queue metrics stay private to the partner.
Industry: Public infrastructure
Challenge: Reviewers requested lineage for a seasonal demand model.
Solution: Experiment Tracking for Governed Teams artifacts were exported with reviewer-friendly diffs.
Outcome: Review cycle shortened by two weeks according to partner PMs; we cite their internal memo, not a public stat.
Industry: Coastal monitoring
Challenge: Jetson boards throttled during August humidity.
Solution: Edge Deployment for Vision Models cohort built rollback plans tied to temperature triggers.
Outcome: Field teams documented two safe derating modes; no uptime guarantees were marketed.