CareCall — AI Companion That Listens Beyond "I'm Fine"

An AI Companion that calls your grandmother daily, earns her trust, and quietly notices what she won't say without ever invading her privacy or breaking her confidence.

Comment

Inspiration

Every time we called our grandmother, she said the same thing: "I'm fine, don't worry about me." But was she really fine?

We discovered this is called protective filtering. Elderly people deliberately hide their pain and loneliness to protect their families from worry. Every elder care app we looked at made this worse by creating surveillance anxiety. Grandma performs wellness instead of expressing it. We asked ourselves: what if instead of monitoring her, we earned her trust first?

That question became CareCall.

What it does

CareCall is a daily AI companion powered by Amazon Nova 2 Sonic that calls your grandmother every morning for a warm, natural 2-minute conversation. It never mentions health monitoring, reports, or family alerts. It just talks like a friend calling to say good morning.

Two layers work simultaneously:

Layer 1 - Trust (Consent) The AI companion named Meera never mentions monitoring or reports. At the end of every call, she simply asks: "Is there anything you'd like me to pass along to Priya today?" Grandma controls what gets shared. Always.

Layer 2 — Truth (Voice Sensing) Even when grandma says "I'm fine", her voice tells a deeper story. CareCall passively tracks:

  • Response latency (slower = fatigue)
  • Filler words like "umm", "huh" (confusion)
  • Conversation length (shorter = low mood)
  • Sentiment patterns over 30 days

Family only ever sees a mood score, never a transcript. Alerts fire only when patterns deviate significantly from her personal baseline.

How we built it

Voice Layer — Amazon Nova 2 Sonic We used Nova 2 Sonic's bidirectional streaming API to build a real-time voice conversation. The system prompt carefully crafts Meera's personality - warm, patient, never clinical. Nova Sonic's prosody awareness detects HOW grandma speaks, not just what she says.

Intelligence Layer — Amazon Comprehend + DynamoDB After every call, the ASR transcript is analyzed by Amazon Comprehend for sentiment scoring. Combined with turn count, filler words, and response patterns, we calculate a mood score. DynamoDB stores a rolling 30-day baseline per user. Anomaly detection compares today's signals against this baseline using statistical deviation.

Alert Layer — AWS SNS When anomaly is detected, AWS SNS sends a privacy-first SMS to the family member. The message contains only the mood observation, never the conversation transcript. Grandma's words stay hers.

Tech Stack:

  • Amazon Nova 2 Sonic (bidirectional streaming)
  • Amazon Comprehend (sentiment analysis)
  • Amazon DynamoDB (session history + baseline)
  • AWS SNS (family alerts via SMS)
  • AWS EventBridge (daily call scheduling)
  • Python + boto3
  • React (family dashboard)

Challenges we ran into

The self-censorship problem Our biggest challenge wasn't technical, it was human. We realized early that if grandma knows the AI reports to her daughter, she will self-censor to keep her daughter happy. We redesigned the entire consent layer around this insight. The AI never mentions monitoring. Trust comes first.

Voice pattern baseline Building a meaningful anomaly detector required at least 3-7 days of call history per user. We solved this by seeding a mock baseline on first use and refining it with real data over time.

Handling real conversations During testing, our "Nani" told Meera she hated Priya, asked her to pass a pav bhaji order, and complained about sentence length all in one call. Meera adapted gracefully to every curveball. That's when we knew the system was truly robust.

Accomplishments that we're proud of

  • Built a voice AI that feels like a friend, not a health monitor
  • Designed a consent layer that respects elderly dignity and autonomy
  • Achieved real anomaly detection using passive voice signals - no invasive questions needed
  • End-to-end pipeline: call → analyze → alert fires as real SMS on family's phone
  • Multilingual ready: English and Hindi supported via Nova 2 Sonic

What we learned

The most important thing we learned has nothing to do with technology.

Grandma says "I'm fine" because she loves you.

Building technology for vulnerable people requires designing for human psychology first — not features. Every technical decision we made was driven by one question: "Does this respect her dignity?"

We also learned that Amazon Nova 2 Sonic is remarkably capable of handling real, messy, unpredictable human conversations not just scripted demos.

What's next for CareCall

  • Twilio integration — real outbound phone calls to any number, no app needed
  • Hindi + regional language — full support for Marathi, Tamil, Telugu
  • Fall detection — if grandma mentions pain or doesn't answer 2 days in a row, escalate to emergency contact
  • Memory layer — remember her birthday, her favorite topics, her grandchildren's names
  • NGO partnerships — deploy for elderly care organizations across India

Built With

Share this project:

Updates