PersonalAngel: an on device multimodal safety investigator with adversarial verification and policy gated action

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

PersonalAngel is an autonomous safety investigator that runs entirely on one edge computer. Seven small pretrainedperception models (object detection, tracking, body pose, zero shot image and text matching, speech recognition withtranslation in 13 languages, sound recognition and multilingual text risk scoring) turn video and audio into timestamped structured events. An evidence graph ranked with personalised PageRank selects the frames shown to a locallyserved vision language model with seven billion parameters. A ReAct master agent with eleven costed tools and acompute budget investigates the primary hypothesis; the same model, prompted with the opposite objective, acts as anadversarial critic whose review is required before any costly action; and a transparent policy gate compares belieftimes harm with the cost of a wrong action, asks the person only when asking is safe, and treats silence as evidence.No raw media, transcript or memory leaves the device. On a real surveillance clip of a bar fight, a desktop workstation with 128 GB of unified memory reached thecorrect escalation in 53.4 s with three model calls. A second case, a real rifle and a gunshot, shows how anevidence floor contains a vision model error; it led to a stricter rule in the current version.