LifeWatch Architecture Diagram
Alert message and follow up message
Querying for historical fall events
Natural engagement

🧠 LifeWatch — AI-Powered Fall Detection & Elderly Care Assistant

💡 Inspiration

We wanted to give families peace of mind and help elders live independently without feeling alone. Falls and medical emergencies often go unnoticed, so we set out to create an AI-powered digital guardian that could see, decide, and respond in real time.

⚙️ What It Does

LifeWatch is an AI caregiver assistant that combines video intelligence and real-time decision-making. It detects falls, monitors wellbeing, and triggers instant alerts through Telegram — keeping elders safe while giving families confidence that help is always near.

🏗️ How We Built It

🧩 System Overview

The fall detection bot leverages AWS Step Functions to orchestrate a workflow of Lambda functions responsible for video processing, fall detection, and real-time notifications. Fall analysis is powered by Claude Sonnet 3.5, with historical fall events retrieved from Bedrock's knowledge base stored in a vector database, powered by Amazon Nova Lite, enabling accurate, low-latency detection in a fully serverless architecture.

🧠 Data Flow

Video Input (Streamlit) Streamlit streams real-time video frames to AWS for analysis.
Trigger Step Function Invokes the ElderlyFallDetectionDAG which sequences detection → classification → notification.
Rekognition Pre-Filter Serves as the first detection layer — checks if a person is lying on the ground and stores relevant images.
Video Invoke Lambda Determines whether a detected instance should be classified as a high-alert fall event using Rekognition outputs.
Fall Agent (Bedrock) Applies AI reasoning to classify fall severity into:

0️⃣ No Fall
1️⃣ Low Alert
2️⃣ High Alert Integrates a Knowledge Base that enables contextual understanding and leverages historical data. The Knowledge Base connects through analysis JSONs produced by the agent invocation Lambda and stored in the events bucket.

Store Events Fall classifications and metadata are stored for future reference.
Telegram Feedback Handler Pushes real-time alerts to subscribed caregivers via Telegram.
Telegram Bot + API Integration Enables interactive caregiver commands:

/subscribe — start receiving alerts
/unsubscribe — stop receiving alerts
/status — check subscription status

The bot supports natural, human-like conversations. You can ask questions such as “How many falls occurred today?”, and it will respond intelligently using data from the Knowledge Base.

🚧 Challenges We Faced

Minimizing false positives from Rekognition without missing real falls.
Building a low-latency AWS pipeline to maintain real-time responsiveness.
Designing Bedrock prompts for consistent, explainable classification.
Handling parallel processing and message delivery across multiple Lambdas.

🏆 Accomplishments

Built a fully agentic workflow capable of accurately classifying fall severity, leveraging historical fall data for contextual understanding.
Built a fully automated end-to-end fall detection system using AWS-native tools.
Combined Rekognition + Bedrock + Step Functions for event reasoning.
Integrated Telegram-based caregiver notifications in real time.
Created a modular, extendable AWS pipeline ready for hospital or home deployment.

📚 What We Learned

The importance of designing and refining a robust architecture workflow for business needs
How to design multi-step AI workflows using AWS Step Functions.
How to use Rekognition as a pre-filter to reduce compute costs from agentic AI.
The importance of event-driven system design for continuous monitoring.
How to integrate external tools with AWS services

🚀 What’s Next

Train a custom vision model which can take in human feedback through the Telegram bot to improve detection accuracy.
Add a verbal check in after an incident is detected to more accurately determine the severity of the incident.
Add wearable sensor inputs such as accelerometer data for incident detection outside of camera range/ in areas requiring more privacy.
Develop a mobile caregiver dashboard with push notifications.
Expand integrations with eldercare partners and clinics.
Strengthen data privacy and PDPA/HIPAA compliance.