Sahay — When Memory Fades, Someone Should Still Be There
I built Sahay, a voice-first AI companion designed to support people living with dementia and their caregivers through personalized guidance, safety monitoring, caregiver alerts, and memory-centered experiences. This is the story of how I combined AI, AWS, voice interaction, and event-driven architecture to build a more human-centered approach to dementia care.
Sahay

"I used several real-world examples to give you a better understanding of the problem and how I addressed it using existing technology. You are gonna find several app-related images for visualization, and you are gonna find some architectural images as well. I used one architecture image in this blog so it doesn't get confusing for readers. For more architectural understanding, you should reference my GitHub repository for this project." ~ Protham
so happy reading, I hope it doesn't bore you, and can be helpful! :)
When memory fades, someone should still be there.
A voice-first AI companion for people living with dementia — and the families caring for them.
Sahay helps caregivers handle difficult moments, supports patients when they become disoriented, and watches for patterns of caregiver exhaustion before they become overwhelming.
Speak naturally. Get help immediately. Keep your loved one safe.
The moment nobody prepares you for
It's 8:30 PM.
Your mother asks the same question for the fifth time.
Then she looks at you and asks:
"Who are you?"
- You know she's confused. You know arguing won't help. But in that moment, you don't know what to say.
You don't need another article to read.
You need someone to tell you what to do — right now.
That's where Sahay comes in.
One voice. Three layers of support.
01 — Help the caregiver
"She's calling me by her sister's name. What should I do?"
- Sahay listens in your language and responds with practical, situation-specific guidance.
- It doesn't give the same generic advice to everyone.
- Sahay remembers the context you provide — relationships, routines, preferences and care stage — and uses that context to make its guidance relevant to your family.
Less searching. Less guessing. More confident caregiving.

02 — Protect the patient
Sometimes the problem isn't a question.
It's that your loved one has taken the wrong bus.
Or missed the expected stop.
Or is standing somewhere unfamiliar asking:
- "Where am I?"
Sahay can monitor an expected route and identify meaningful deviations.
When a high-confidence anomaly occurs, Sahay can:
- Detect → Respond → Alert
- It generates a calming, personalized voice message for the patient while simultaneously alerting the caregiver with relevant context.
- The goal isn't to control the patient.
It's to help them get safely back to someone they trust.

03 — Look after the person doing the caring
Caregiver burnout rarely happens in one conversation.
- It builds quietly.
"A bad night.
Another exhausting day.
Another week of feeling like you're doing everything alone."
- Sahay tracks changes across conversations instead of reacting to one difficult moment.
- When it sees a sustained decline, it can nudge the caregiver toward real human support and relevant resources.
Because the caregiver needs care too.
What makes Sahay different?
It's not another medical information app.
Traditional apps often start with:
Search → Read → Understand → Act
But during a crisis, that's not how people behave.
Sahay is designed around:
- Speak → Understand → Guide → Escalate when necessary
The interaction is designed around something people already know how to do:
talk.
Personalization that actually matters
Imagine saying:
"She keeps calling me her sister's name."
A generic chatbot might explain why people with dementia confuse relationships.
Sahay can use the patient's stored context to understand who that sister is, what relationship they have, and what guidance is appropriate for that individual.
- This is powered by retrieval-augmented generation, so patient context isn't simply buried inside one giant prompt.
- The AI retrieves the right context when it is needed.
- The system can also adapt its guidance based on the patient's care stage, making the response more appropriate to the situation instead of treating every family the same way.
From one AI model to a coordinated safety system
Sahay doesn't rely on one giant prompt to handle everything.
Behind the scenes, specialized agents work together.

"Does someone else need to be contacted?"
- This orchestration runs through AWS Step Functions, with confidence gates around autonomous actions.
- That means Sahay is designed to understand an important difference:
"I can help."
versus
"A human needs to be involved."
When location becomes part of the conversation
A patient's words alone don't always tell the whole story.
Neither does their location.
Sahay combines contextual signals to reduce unnecessary intervention.
For example:
Patient asks:
- "Where am I?"
AND
- The patient's route has significantly deviated from the expected journey.
That combination can trigger a stronger intervention than either signal alone.
- The system can generate a personalized calming response and notify the caregiver with the relevant context.
- This approach can also be extended to other safety signals such as expected destinations, safe zones and unusual movement patterns.
Context turns a signal into a meaningful action.

A safety net anyone can activate
What if the patient can't operate the phone?
Sahay includes a simple "I'm lost / Help" experience and a caregiver-generated QR card.
A bystander can scan the QR code and access the appropriate assistance flow and caregiver contact information.
No account creation.
No complicated interface.
Just a simple path from:
"I found someone who needs help."
to
"Their caregiver has been contacted."
Memory can be more than data.
Sahay's Memory Theater transforms familiar family photographs into a calming, interactive experience.
Photos of:
- Family. Places. People. Moments.
- can be presented as a narrated visual sequence.
Instead of asking a confused patient to remember "where they are",
- Sahay brings familiar memories to them.
- The experience is designed to make technology feel less like a clinical tool and more like something familiar and comforting.

One system for the patient and the caregiver
Sahay isn't only about what happens during an emergency.
The same context gathered during everyday interactions can help build a clearer picture for the caregiver.
Conversations, distress signals, route events and important patient context can be brought together so caregivers don't have to reconstruct what happened from scattered interactions.
This creates the foundation for features such as:
- Caregiver summaries of important interactions
- Recurring confusion and distress patterns
- Routine and destination monitoring
- Safe-zone based alerts
- More personalized daily reminders
- Multilingual voice interaction
- Additional safety signals from connected devices in the future
The idea is simple:
Don't make the caregiver collect all the information themselves.
Let the system organize the signals and surface what actually matters.
Built for India
India's dementia burden is growing rapidly, while many families still rely primarily on informal caregiving.
The problem isn't simply access to information.
It's access to usable support.
Sahay is designed around that reality:
- Voice-first
- Talk instead of navigating complicated interfaces.
- Vernacular-ready
- Designed for interaction in local languages.
- Low-friction
- Useful even when the caregiver has almost no time.
- Human escalation
- AI assists; people remain in the loop when situations become serious.
- Privacy-aware
- Sensitive audio processing is designed around short-lived processing rather than continuous cloud recording.
Built on AWS. Designed around real-world events.
Sahay isn't just a chatbot with an AWS logo attached.
The deployed system uses AWS services according to the job they need to perform.
- Amazon API Gateway Connects the application to backend services and provides the API layer.
- AWS Lambda Runs application logic and event handlers without requiring dedicated servers.
- Amazon DynamoDB Stores operational data such as patient profiles, conversations, distress signals and route configurations.
- Amazon Bedrock Powers the AI reasoning and contextual intelligence behind the system.
- Amazon Bedrock Knowledge Bases / OpenSearch Serverless Support retrieval of personalized patient context for RAG-based responses.
- AWS Step Functions Orchestrates the Triage, Care-Guidance and Escalation workflows.
- Amazon EventBridge Drives scheduled monitoring such as burnout-trend checks and other event-driven workflows.
- Amazon SNS Provides caregiver and escalation notifications.
- Amazon Polly / Amazon Transcribe Enable voice interaction and spoken responses.
- Amazon S3 Handles application assets and short-lived audio-related data.
- Amazon Cognito Provides authentication for the caregiver experience.
- Amazon Amplify Hosts and delivers the live application.
- Amazon CloudWatch / AWS X-Ray Provide monitoring, logs, metrics and distributed tracing across the system.
The result is an architecture where a new event can flow through:
- Signal → Context → AI Reasoning → Confidence Gate → Action → Human Escalation
That makes Sahay more than a chatbot.
It's an event-driven safety system.

Designed with guardrails, not just intelligence.
AI should not silently make high-impact decisions.
That's why Sahay follows three principles:
Confidence-gated actions
High-confidence signals can trigger automated flows. Lower-confidence findings can request confirmation rather than acting blindly.
Human escalation
Sahay assists caregivers; it does not replace doctors, caregivers or emergency services.
Fail safely
If an AI service fails, the system should fall back to a safe response rather than leaving a caregiver without guidance.
Privacy by design
Sensitive audio should not become a permanent record simply because an AI system processed it. Sahay is designed around short-lived processing wherever cloud audio processing is required.
Built to respond in seconds, not minutes.
For a safety companion, latency matters.
In the current architecture, the voice interaction is designed around a fast end-to-end flow:
Speak → Transcribe → Understand → Generate → Speak back
Metric | Your target | Assessment |
|---|---|---|
Page startup | ~300ms P50 | Excellent |
STT | ~500ms P50 | Very good |
AI response | ~1.5s P50 | Good |
TTS | ~200ms P50 | Excellent |
Speak → hear reply | ~2.5s P50 | Very good |
Speak → hear reply | ~4s P95 | Good |
Speak → hear reply | ~7s P99 | Acceptable for an AI voice workflow |
Caregiver alert | ~400ms P50 | Very good |
Caregiver alerts are designed to reach the configured destination within approximately sub-second to low-second latency, depending on the external notification path.
Sahay also uses alert deduplication so that one ongoing incident does not repeatedly spam the caregiver.
These aren't just performance numbers.
In a stressful situation, every unnecessary second and every unnecessary alert matters.
The bigger idea
Sahay is built around a simple idea:
Technology should not replace human care. It should make human care easier.
Dementia care is not a single problem that can be solved with one chatbot, one alert, or one medical database. It involves communication, safety, routines, memory, caregiver wellbeing, and countless small decisions made throughout the day.
Sahay brings these pieces together into one support layer.
It listens when communication is difficult.
It provides context when decisions are unclear.
It watches for meaningful safety signals.
It connects caregivers when human intervention is needed.
And it helps preserve familiarity through memories and personal context.
The long-term vision is to build technology that quietly works around the patient and caregiver, rather than forcing them to constantly work around the technology.
AI handles the signals.
The system handles the coordination.
People remain at the center.
That is the bigger idea behind Sahay.
Thank you for reading, It means a lot~
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article