Inspiration

The word Saathi (साथी) means companion in Hindi and Marathi. We wanted to build a companion that helps older adults living with memory loss stay connected with their families, maintain their independence, and feel cared for even when their loved ones live far away.

Imagine a 78-year-old grandmother living alone in Pune, speaking Marathi, while her granddaughter lives in Toronto and primarily speaks English. Add memory loss, hearing difficulties, and the everyday challenge of remembering medication, and staying connected becomes much harder.

This inspired us to build Saathi-Alexa: an AI-powered caregiving companion that bridges language barriers, supports daily routines, preserves family memories, and connects caregivers when help is needed. Our goal was to make technology more human, accessible, and compassionate.

What it does

Saathi-Alexa is an Alexa+ integration powered by a self-hosted Model Context Protocol (MCP) server. It connects conversational AI with caregiving tools and a Family Circle application.

Key features include:

Multilingual conversations: Supports eight languages, helping users interact in a language they understand.
Memory assistance: Gently answers questions about the date, daily routines, family members, and personal information without quizzing or correcting the user.
Medication management: Organizes medication schedules by colour and purpose, records doses, prevents duplicate dose confirmations, and tracks remaining supplies.
Intelligent check-ins: Analyzes emotional well-being and screens for safety concerns. Loneliness prompts a gentle family notification rather than an unnecessary emergency alarm.
Emergency escalation: Uses AWS Step Functions to coordinate responses to potential emergencies. The workflow contacts family first, then a neighbour, and progresses to emergency guidance if nobody responds. Acknowledging an alert stops further escalation.
Translated family communication: Allows family members to exchange messages across English, Hindi, and Marathi, preserving the sender's intended words across language barriers.
Family Memory Book: Saves personal stories, translates them into family members' languages, and uses Amazon Polly to narrate them.
Care coordination: Supports doctor appointments, caregiver assignments, medication refills, voice notes, and weekly family summaries.
Accessibility: Provides voice interactions, captions, large text, high contrast, and slower speech for users with different accessibility needs.
Ring doorbell integration: Includes a simulated Ring doorbell and webhook-based event handling to support caregiving workflows.

Saathi is designed to support caregivers rather than replace them, keeping human connection at the centre of the experience.

How we built it

We built Saathi-Alexa using JavaScript, Node.js, Alexa+, MCP, and an AWS-based architecture.

The core system is a self-hosted MCP server that exposes caregiving capabilities through tools, resources, and prompts. It uses Streamable HTTP and implements MCP protocol features, including OAuth 2.1 authorization and scope-based access control.

Our architecture combines several components:

AI and conversational intelligence: Claude through Amazon Bedrock or the Claude API powers the simulated Alexa+ experience and tool selection.
AWS Lambda and API Gateway: Provide a serverless deployment path for the MCP server and Ring event handling.
Amazon DynamoDB and Amazon S3: Support application data and memory-book audio storage.
Amazon Translate and Amazon Comprehend: Support multilingual communication and sentiment analysis.
Amazon Polly: Generates narrated audio for saved family stories.
Amazon EventBridge and AWS Step Functions: Coordinate events and escalation workflows.
Amazon SNS: Supports caregiver notifications.
Amazon Cognito: Provides OAuth 2.1 authorization for the MCP integration.
AWS CDK: Defines the cloud infrastructure as code.
Family Circle web app: Gives family members a dashboard for alerts, mood trends, messages, memory stories, and caregiving coordination.

One of our most important architectural decisions was to make the project runnable without a cloud account. Each AWS integration sits behind an adapter with a local alternative. The same escalation state-machine definition can run through AWS Step Functions or a built-in offline interpreter, allowing us to test the workflow locally.

The project can be started locally with npm install and npm start, using an offline planner when no AI model is configured.

Challenges we ran into

One major challenge was coordinating multiple services while keeping the experience reliable for people who may depend on it. Conversational AI, medication records, multilingual processing, notifications, and emergency workflows all have different requirements.

Making safety multilingual was another significant challenge. Sentiment analysis alone is not enough to detect emergencies, especially when a user speaks Hindi or Marathi. We implemented a multilingual phrase-based safety screen alongside sentiment analysis, including support for romanized phrases. Unsupported languages can be translated into English before sentiment analysis.

We also had to prevent duplicate emergency workflows. AI assistants may retry tool calls, and a frightened person might ask for help more than once. Saathi is designed to join an existing open alert rather than start multiple escalations for the same emergency.

Another challenge was distinguishing genuine danger from harmless language. For example, the word "fall" may refer to falling prices or falling leaves rather than a person falling. We developed multilingual safety evaluations to identify false alarms and missed emergencies.

Finally, we focused on dignity and privacy. Ordinary sadness should not automatically trigger an emergency response, and messages should not leave the home without confirmation. Balancing helpful automation with personal choice was central to our design.

We also recognize the limitations of our current implementation: the Ring device is simulated, the offline sentiment model is a lightweight alternative, and the AWS stack has not yet been deployed to a live AWS account.

Accomplishments that we're proud of

We are proud to have developed a comprehensive caregiving system rather than just a conversational chatbot.

Built an MCP server with 16 tools, three resources, a resource template, and two prompts.
Designed an end-to-end escalation workflow with family acknowledgement and controlled progression to further assistance.
Created a Family Circle application for coordinating caregiving across family members who speak different languages.
Implemented medication tracking, refill reminders, appointment coordination, voice notes, and a translated memory book.
Designed a multilingual safety evaluation suite with 44 safety cases and 25 tool-routing cases.
Built local alternatives for AWS integrations so the project can run and be tested without a cloud account.
Developed AWS infrastructure as code covering 65 resources across 15 AWS services.
Added automated tests, deployment checks, and accessibility audits to strengthen reliability.

We are particularly proud of the design principle behind Saathi: low mood deserves compassion, while genuine danger deserves a coordinated response. The system is designed to send a gentle nudge for loneliness rather than treating every emotional check-in as an emergency.

What we learned

We learned that building AI for caregiving requires much more than connecting a language model to a set of tools.

MCP taught us how to expose structured capabilities to an AI assistant while keeping the system modular. AWS Step Functions helped us design explicit, testable escalation workflows. Building local adapters taught us the value of separating application logic from cloud infrastructure.

We also learned that multilingual AI requires careful engineering. Translation and sentiment analysis have limitations, so safety-critical decisions need additional safeguards and systematic evaluation.

Most importantly, we learned that accessibility, privacy, reliability, and dignity must be built into the architecture from the beginning. A caregiving assistant should never make people feel tested, exposed, or unnecessarily alarmed. It should offer support while keeping families and human caregivers involved.

What's next for Saathi-Alexa

Our next goal is to move Saathi-Alexa from a locally testable prototype toward a thoroughly validated, real-world caregiving solution.

Our roadmap includes:

Live AWS deployment: Deploy the infrastructure and validate the complete system against real AWS services using end-to-end smoke tests.
Real device integration: Move beyond the simulated Ring doorbell and validate supported integrations with actual devices.
Stronger safety evaluations: Expand multilingual test cases using feedback from caregivers, native speakers, and relevant healthcare professionals.
Improved language quality: Have native speakers review translations, spoken responses, and culturally appropriate phrasing.
Enhanced privacy and security: Develop stronger authentication for the Family Circle app, household isolation, and clearer consent controls.
More personalized assistance: Adapt check-ins, routines, accessibility settings, and communication preferences to individual needs.
Real-world user testing: Work with older adults, families, and caregivers to evaluate usability, reliability, and trust.
More robust emergency handling: Test failure scenarios, notification delivery, escalation timing, and recovery from interrupted workflows.

Our long-term vision is to make Saathi a dependable, inclusive companion that helps older adults maintain independence while giving families greater peace of mind.

Saathi means companion, and we want to build technology that truly lives up to that name: not replacing human care, but bringing people closer together when distance, memory loss, or language barriers get in the way.