International Journal of Scientific Research and Engineering Development

Emergency response applications overwhelmingly rely on a victim's ability to manually trigger a distress signal, an assumption that fails precisely in the scenarios where intervention matters most: when a person is unconscious, physically restrained, under active assault, or otherwise unable to reach a device. Prior work has advanced human activity recognition, speech emotion recognition, audio event detection, and keyword spotting largely as isolated, single-modality problems, leaving open the question of how their heterogeneous, individually unreliable outputs can be combined into a single, actionable, and privacy-preserving judgment. This paper addresses that gap by presenting Javipra AI, an autonomous, privacy-preserving emergency detection framework whose central contribution is the Javipra Confidence Fusion Engine (JCFE), a confidence-weighted multimodal sensor fusion mechanism that departs from conventional weighted-average fusion by explicitly modeling inter-module agreement and temporal persistence as first-class evidentiary signals, rather than treating every module's confidence as independently and identically informative. Rather than depending on a single modality or a manually activated switch, the framework continuously and locally analyzes microphone audio, accelerometer and gyroscope streams, and GPS context through five independent artificial intelligence modules, voice keyword spotting, speech emotion recognition, audio event detection, human activity recognition, and GPS context analysis, each contributing a calibrated confidence score. JCFE combines these scores with contextual risk weighting into a unified emergency probability, which an adaptive decision engine compares against a context- and history-sensitive threshold explicitly designed to suppress false alarms without sacrificing sensitivity to genuine danger. On confirmation of an emergency, the system autonomously shares live location, notifies pre-designated trusted contacts, records encrypted emergency audio, and continues heightened monitoring, entirely without requiring deliberate user action. The complete inference pipeline is designed to execute on-device using TensorFlow Lite, so that raw audio, motion, and location data need never leave the device except as part of a minimized, authenticated emergency alert, giving the framework a privacy posture that is structural rather than policy-based, while enabling low-latency, offline-capable, battery-conscious operation on commodity smartphone hardware. This paper details the system architecture, the mathematical formulation of the confidence, risk, fusion, and decision functions, the algorithmic realization of JCFE, a proposed implementation stack, and a staged experimental methodology, including ablation studies and baseline comparisons, intended to validate the framework's detection accuracy and false-alarm behavior prior to field deployment. By reframing personal safety as an autonomous multimodal inference problem rather than a manually triggered event, Javipra AI is positioned to extend meaningful protection to precisely the population of emergencies, assault, abduction, medical incapacitation, falls, and restraint, that existing manual-activation systems structurally cannot address

Jahnavi Somaraju, Chethan Kumar Reddy P, "Javipra AI: A Confidence-Weighted Multimodal Sensor Fusion Framework for Autonomous, Privacy-Preserving Emergency Detection and Response on Edge Devices" International Journal of Scientific Research and Engineering Development, V9(4): Page(418-436) May-June 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.