CareBridge: An Autonomous Agentic Elder Companion on WhatsApp

Inspiration

Every adult child living away from their aging parents shares the same quiet anxiety:

Did Papa take his blood pressure medicine this morning?
Did Ma sleep well, or is that knee pain keeping her awake again?
Was that mild dizziness she mentioned just fatigue, or an early warning sign?

Over 140 million seniors live in India alone, with more than 40% living alone or hundreds of miles away from their working children. Yet, almost every digital senior healthcare solution fails with an app abandonment rate exceeding 90%. Elderly parents refuse to download new apps, forget login passwords, and find modern multi-level navigation overwhelming.

Their sole digital comfort zone is WhatsApp.

When we saw the challenge for the AWS CDS Agentic AI Partner Hackathon, our team asked: What if we stopped treating communication services as dumb notification pipes for OTPs and delivery alerts? What if AWS CDS (WhatsApp) became the proactive, empathetic, agentic communications backbone for eldercare?

That was the birth of CareBridge — an autonomous AI companion on WhatsApp that proactively checks in on elders, tracks medication adherence through natural vernacular conversation, and autonomously escalates emergencies to families in seconds using Amazon Bedrock AgentCore.

What It Does

CareBridge operates with zero app friction for elders and provides a calm, dignified peace-of-mind dashboard for adult children:

Autonomous Daily Check-ins on Native WhatsApp: At scheduled hours (e.g., 9:00 AM), CareBridge reaches out proactively over WhatsApp. It greets elders respectfully in colloquial Hindi, Hinglish, or English (using "Aap"), asking about their morning walk, breakfast, sleep quality, and prescribed medicines (e.g., Amlodipine 5mg).
Conversational Extraction of Adherence: Instead of forcing seniors to check checkboxes, CareBridge understands natural vernacular replies (e.g., "Namaste beta, maine chai pee li aur dawai bhi le li"), extracting structured clinical records (medications taken, meals eaten, mood score 1–5).
Autonomous Emergency Escalation via Bedrock AgentCore: If an elder mentions acute distress (e.g., "My chest feels tight", "I fell down near the bathroom", "I feel very dizzy"), Amazon Bedrock autonomously halts routine chatter, invokes the alert_family(severity, summary) tool, dispatches local emergency numbers (112 and 108), and fires priority WhatsApp alert templates to family members.
Cascading Multi-Tier Family Alerts: If the primary family contact does not acknowledge the high-severity alert within 15 minutes, the system autonomously cascades the escalation to secondary family members.
Calm, Accessible Family Dashboard: Hosted on Amazon CloudFront and API Gateway, the React SPA provides 7-day and 30-day medication adherence charts, mood trajectories, searchable call/chat transcripts, and an interactive 6-Step Onboarding Wizard to add new elders dynamically.
Privacy by Design: Strict double opt-in consent via WhatsApp, ephemeral audio processing with zero voice storage, and configurable DynamoDB TTL auto-purge (30, 90, or 365 days).

How We Built It

CareBridge is built on a 100% serverless, cloud-native architecture deployed via AWS CDK (TypeScript):

[Elder on WhatsApp]
        ↕ (Text / Audio Notes / Voice)
[AWS End User Messaging Social] (Meta WABA in ap-south-1)
        ↓ (Inbound Webhook)
[Amazon SNS A2P Topic] (Cross-Region Event Ingestion)
        ↓
[CareBridge Router Lambda] (Python 3.12 ARM64 in us-east-1)
        ↕
[Amazon Bedrock AgentCore] (Nova Micro / Nova 2 Sonic + Converse Tool Use)
        ↕
[Amazon DynamoDB] (10 Partitioned Tables with TTL Purge)
        ↕
[Amazon API Gateway & CloudFront] ↔ [React SPA Family Dashboard]

AWS End User Messaging Social (ap-south-1): Serves as our primary communications backbone, handling high-throughput bi-directional WhatsApp messaging with Meta Cloud API compliance.
Amazon SNS (A2P): Provides a resilient cross-region fanout pipeline, decoupling the Mumbai phone endpoint from our North Virginia backend with deduplication and idempotency.
Amazon Bedrock AgentCore (amazon.nova-micro-v1:0 & amazon.nova-sonic-v1:0): Powers empathetic conversational reasoning with sub-second response times. Using the Bedrock Converse API with tool configuration, the model autonomously decides when to trigger alert_family versus continuing friendly dialogue.
AWS Lambda (Python 3.12, ARM64): Micro-services for message routing, schedule cron execution, voice audio handling, and REST APIs.
Amazon DynamoDB: 10 single-digit latency tables partitioned by elder_id, phone, and call_id, with Point-in-Time Recovery (PITR) and automatic TTL expiration.
Amazon CloudFront & S3: Delivers a WCAG 2.1 AA accessible React SPA with warm palettes and instant live updates.

Unit Economics & Cost Efficiency

Because CareBridge is built serverless with zero idle compute costs, the unit economics are exceptionally scalable:

The total monthly cost per elder $\text{Cost}_{\text{elder}}$ is defined by:

$$\text{Cost}{\text{elder}} = C{\text{CDS}} + C_{\text{Bedrock}} + C_{\text{Compute}} + C_{\text{Storage}}$$

Where for a daily 3-minute conversational check-in (30 days/month):

AWS CDS Social Messaging ($C_{\text{CDS}}$): 30 conversation sessions $\times$ $\$0.045 \approx \$1.35$
Amazon Bedrock Nova Micro ($C_{\text{Bedrock}}$): $30 \text{ turns} \times 1,500 \text{ tokens} \times \frac{\$0.00035}{1,000} \approx \$0.16$
AWS Lambda & API Gateway ($C_{\text{Compute}}$): $60 \text{ invocations} \times 512\text{MB} \approx \$0.04$
DynamoDB & S3 ($C_{\text{Storage}}$): Pay-per-request storage with TTL $\approx \$0.30$

$$\text{Cost}_{\text{elder}} \approx \$1.85 \text{ per elder / month } (\approx ₹155 / \text{month})$$

At a family subscription price of $\$9.99 / \text{month}$, CareBridge generates a $78\%+$ gross margin, proving that empathetic eldercare can be both life-saving and economically sustainable.

Challenges We Faced

Cross-Region CDS Webhook Ingestion: Our verified WhatsApp business phone number was provisioned in ap-south-1 (Mumbai), while our Bedrock models and DynamoDB stack were optimized in us-east-1 (N. Virginia). We engineered a cross-region SNS subscription with granular IAM resource policies (AllowSnsFromApSouth1), achieving seamless intercontinental delivery in under 80ms.
Deterministic Safety Nets vs. Generative Warmth: LLMs can hallucinate or downplay medical emergencies. We addressed this with a dual-gate safety architecture: a deterministic regex/keyword pre-filter and post-filter that wraps the Bedrock AgentCore. If an elder reports severe pain, the system forces injection of emergency numbers (112 and 108) and guarantees immediate family alert delivery regardless of model output.
Bedrock Chain-of-Thought Tag Leakage: During real WhatsApp tests, reasoning models generated internal thoughts enclosed in <thinking>...</thinking> tags (e.g. "Papa took his medicine, no alert needed"), which initially leaked into user chats and contaminated conversation history. We implemented robust regex sanitation (clean_reply_text) and modified history builders to enforce strict alternating turns (user $\leftrightarrow$ assistant) per Bedrock Converse API rules.
Strict E.164 Formatting in AWS End User Messaging Social: AWS End User Messaging Social strictly requires phone numbers to include the leading + prefix (e.g. +917842359784). Standardizing normalizers across the pipeline resolved InvalidParametersException errors.

What We Learned

CDS as an Agentic Backbone: Communication services shouldn't just be passive pipelines. When integrated with AgentCore tool execution, WhatsApp becomes an interactive sensory organ and voice for autonomous agents.
Atomic Tools Beat Long System Prompts: Rather than telling an LLM "write an alert if someone is sick", providing a strict atomic tool alert_family(severity, summary) gives predictable, auditable, and testable behavior.
Accessibility is Compassion: High-contrast fonts, calm off-white aesthetics, zero required logins for seniors, and native vernacular language are essential for senior technology adoption.

What's Next for CareBridge

Multilingual Expansion: Expanding beyond Hindi and English to Gujarati, Marathi, Tamil, and Telugu using Amazon Nova 2 Sonic.
Geriatric Clinic & EHR Sync: Bidirectional integration with hospital portals (FHIR/HL7) so doctors can verify real-world medication adherence.
Wearable Fall Detection Integration: Syncing heart-rate spikes and fall sensors from smartwatches directly into the CareBridge distress net.
Interactive WhatsApp WebRTC Calls: Scaling low-latency voice streaming via Amazon Bedrock Nova Sonic directly