Silent-Sentinel
‘SilentSentinel’-Cries never go unnoticed
Created on 30th December 2025
•
Silent-Sentinel
‘SilentSentinel’-Cries never go unnoticed
The problem Silent-Sentinel solves
The Problem It Solves
Unnoticed emergencies at home:
Many critical incidents such as falls, medical collapses, distress episodes, or environmental dangers occur when a person is alone, disoriented, or physically unable to call for help.
Dependence on user action:
Most existing safety systems rely on wearables, panic buttons, or phone interactions, which fail when the user is unconscious, confused, asleep, or unable to reach a device.
Privacy-invasive monitoring:
Camera-based solutions raise serious privacy concerns and are often rejected by elderly users and families.
Delayed response in real-world scenarios:
Even a few minutes of delay in detecting an emergency can significantly worsen outcomes, especially in cases of strokes, falls, or internal injuries.
What Makes Silent Sentinel Different
Passive, always-on listening (no wearables):
Silent Sentinel works without requiring the user to wear devices or press buttons. It continuously monitors the environment for critical audio cues.
Sound-based intelligence, not cameras:
The system relies entirely on audio signals, preserving privacy while remaining effective in real-world home environments.
Multi-layered detection instead of raw confidence:
Unlike traditional models that trigger alerts based on a single confidence score, Silent Sentinel combines:
ML-based sound detection (YAMNet + custom models)
This was our self trained custom gunshot model with 98.6% of accuracy.
Temporal and pattern-based analysis
Contextual reasoning via an LLM before escalating emergencies
Voice-initiated emergency support (“SS Help”):
If sound detection fails or isn’t applicable, users can directly trigger an emergency using predefined voice phrases, ensuring no cry for help goes unnoticed.
Action-driven AI via MCP:
The system doesn’t just detect—it acts. AI decisions trigger real-world actions like WhatsApp alerts through a structured MCP workflow, rather than hardcoded scripts.
Designed for real homes, not controlled labs:
Silent Sentinel is built with real-world noise, false positives, and edge cases in mind—making it practical, resilient, and trustworthy.
Challenges we ran into
Challenges We Ran Into
False positives in real-world audio detection
Models that showed high accuracy on datasets behaved unpredictably in real homes. Everyday sounds like door slams or object drops were occasionally misclassified as emergencies, forcing us to rethink purely confidence-based triggers.
Designing a multi-layered verification system
To reduce false alarms without missing real emergencies, we built a tiered pipeline combining YAMNet outputs, a custom gunshot model, temporal sound patterns, and contextual reasoning via an LLM.
Custom model integration mismatches
Our gunshot model initially returned incorrect probabilities due to a mismatch between sorted YAMNet outputs and index-based model expectations. Identifying and fixing this required a deep understanding of audio model data formats.
MCP integration and tool execution
Defining MCP tools was not enough—initially, they were never triggered. The real challenge was wiring the reasoning engine to actionable tools by correctly exposing capabilities and passing tool context into the LLM.
Alert delivery constraints in real-world systems
SMS delivery posed verification and regional limitations during the hackathon. We pivoted to WhatsApp alerts to ensure reliable, instant notifications without blocking the end-to-end emergency workflow.
Audio session and lifecycle issues
Browser audio contexts frequently closed or failed to resume, requiring careful session management and error handling to keep detection stable over long monitoring periods.
Preventing duplicate or repeated alerts
Ensuring a single emergency didn’t spam caregivers required cooldown logic, session tracking, and clear user controls for resuming or pausing monitoring.
Tracks Applied (2)
Best Innovation
AWS
AWS