Safe House Vision & Response

Safe House Vision & Response is a local-first home security system: a routing cascade for existing home cameras. A free local detector discards almost everything; only candidate events rise to a hosted verifier (Muse Spark Open, via Baseten); only verified events become a graduated WhatsApp alert, with the evidence preserved under SHA-256 chain of custody. A human always decides — the system never accuses anyone on a detector's word alone. How it uses Meta AI: Muse Spark Open is the verification stage of the cascade (low-effort calls with a healthy output limit). In loop v1 (Oct 31) it judged 6 frames of an authorized test clip and returned observable facts with confidences per frame: possible package theft 0.62 at 4 s, suspicious behavior 0.85 at 6 s, package delivery 0.95 at 10 s. The graduated alert was received on WhatsApp at 12:13 PT; the clip hash (D815…D330) is linked to the run hash (194F…0253); total egress was ~7.3 MB to a single destination (inference.baseten.co), measured, not asserted. The honest finding: the model over-labelled a simulated delivery as possible theft while its observable facts were correct — the live proof of why the architecture requires a verifier and a human. Declared limits: in v1 the alert is composed and sent manually; automating it is the next iteration. Single authorized test clip; frame timestamps are approximate (sampling interval). The system never classifies a person as dangerous by appearance, clothing, or belongings — observable facts only, with uncertainty declared.