What it is
Near-miss decisioning for construction video, with Jev on structured tracks.
What it does: Turns short behavior summaries into typed hazard judgments and recommended actions.
Why it belongs in the showcase: Concrete Jev 1.14 + Box demo with offline fallback and large documented pipeline.
Caveats: AGPL; experimental images; privacy review required.
Evidence
Public AGPL-3.0 Python app (0★ at intake, size ~19305) created 2026-10-16. README (Korean/EN detail) pins model jev-1.14.0 via POST /v1/systemone, documents TYPESAFE_API_KEY, offline stub mode, demo scenarios with images, and notes image/Box as TypeSafe Dreamer-ready preview. YOLO text summaries are the default state; ROI crops go out only on image fallback.
Caveats
- AGPL-3.0 (Ultralytics YOLO) — sharing the web UI may force source disclosure
- Image/Box path is experimental per TypeSafe; not a full vision model
- Site privacy/PII rules still apply to CCTV frames
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.