Saathi

saathi full story
pattern detection
family approval
care dashboard

Inspiration

Millions of Indian parents live alone while their children work in other cities. The children call every day, but a phone call can't tell you that Maa felt dizzy three times this week, in the same week she missed her blood-pressure tablet three nights. Each conversation sounds fine on its own; the danger only shows up across days. We wanted Alexa to be the one who notices, and to bring the family in before a small pattern becomes an emergency.

What it does

Kamla, 72, lives alone in Jaipur. Her son Rohan works in Bengaluru. Kamla talks to Alexa the way she'd talk to a person, and the Saathi add-on ("saathi" means companion in Hindi):

Logs medicines from how people describe them. "I took my BP tablet" becomes Amlodipine, night dose.
Remembers across sessions. A care-signal engine scores the last 7 days and explains why in plain words.
Spots risky combinations. Repeated dizziness plus missed blood-pressure doses leads Saathi to suggest a doctor's home visit.
Keeps the family in control. Every booking waits for approval. Rohan asks "How is Maa doing?", hears a briefing and says "Approve it", and then Kamla hears the visit is booked.
Puts emergencies first. A fall, chest pain or breathlessness gets an immediate "call 112" and an urgent alert to every family member.
Never diagnoses. It notices patterns and brings in people.

How we built it

Saathi MCP server: Python with the official MCP Python SDK v2, over Streamable HTTP, serving 2025-11-25-era clients. It has 8 tools (5 for the parent, 3 for the family), and each returns a say line written for voice plus structured data.
Memory and care-signal engine: SQLite plus deterministic rules over a 7-day window, so Saathi remembers across sessions. There's no LLM inside the server, so it's fast and predictable: the slowest tool call measured 205 ms over a public Cloudflare tunnel (27 ms locally), within Alexa+'s 500 ms requirement.
Approval-gated care requests: requested, then awaiting family approval, then booked or declined. The booking partner is simulated for the demo.
Simulated Alexa+ experience: an Echo Show in Kamla's home and Rohan's phone, side by side in the browser, with real speech in (Web Speech API) and spoken replies out. An LLM agent (Groq, tool calling) picks the Saathi tools and calls the real MCP server. Every tool call and its latency is shown on screen, and a live care dashboard updates after every turn.
Open source: we extracted what we learned into mcp-voice-kit, a new MIT-licensed library for MCP servers whose answers are spoken aloud, with 32 tests.

Challenges we ran into

Reaching production Alexa+. The Alexa AI CLI installs from a private AWS CodeArtifact registry that needs AWS credentials. As students without an AWS account (and with hackathon credits already exhausted), we couldn't install it. So we took the simulated Alexa+ path the rules allow, kept the MCP server real and spec-compliant, and documented the blocker in our friction log.
MCP SDK v2 changes. FastMCP was renamed to MCPServer, transport settings moved, and DNS-rebinding protection rejected tunnelled requests until we rewrote the Host header.
Designing for voice. Tool output that looks fine on screen (markdown, ISO dates, long IDs) sounds terrible when read aloud.

Accomplishments that we're proud of

A complete, believable care loop (parent, pattern, family approval, booking, reassurance) that runs end to end with every MCP call under 500 ms.
Saathi explains why it's worried in words a family understands, instead of just showing a risk score.
Understanding everyday Hindi words like chakkar (dizziness), because that's how parents actually talk.
A reusable open-source library that came out of real problems we hit.

What we learned

Voice-first MCP design is its own discipline: tool results must be speakable, IDs and dates must sound natural, latency is a feature, and anything involving health or money needs a human in the loop.

What's next for Saathi

A live Alexa+ add-on with OAuth 2.1 account linking, so every family gets their own Saathi
Real booking partners (home lab tests, doctor visits, pharmacies) behind the same approval flow
Proactive evening check-ins and a weekly summary for the family
Hindi and other Indian languages