Alongside

Inspiration

Mary is eighty-one and lives alone. She is sharp and independent, and she still gets stuck — not because she has forgotten how to make breakfast, but because the morning has a lot of small steps and nobody is there to say which one is next. Her daughter Anna has explained it a hundred times.

Alongside is for the hundred-and-first.

What it does

Anna describes a routine once, in her own words, and attaches the photo that makes it obvious which container she means. AI turns that into a structured proposal — time, days, steps, what to say if something is not where it should be — and asks about what she did not say rather than inventing it: "What preparation instructions should I give Mary?" She reviews exactly what the assistant will say and what the television will show, then publishes. Nothing runs until she does.

Next morning Mary's television shows one thing: a greeting, Anna's photo, one instruction, and a few large buttons. She can use the remote or just talk. She can ask what she's having, ask for the next step, say "give me fifteen minutes", say "the container isn't there" — and get Anna's own approved answer about the packets in the pantry — or ask for help, which appears on Anna's dashboard immediately with the activity and step attached.

If she wanders off and comes back, Alongside says "You were on step two. Would you like to continue?" Because showing an instruction is not the same as finishing it, and this product only counts what Mary actually confirmed.

How we built it

One Node backend owns the truth. SQLite separates a routine template from each day's occurrence, which is what makes "postpone this morning only" and "you were on step two" possible at all. Server-side scheduling, Server-Sent Events for browsers and a WebSocket for the television, idempotency keys so a repeated speech result cannot act twice, and version checks so a stale screen cannot overwrite a newer one.

The Fire TV app is React Native built to a real Fire OS APK — leanback launcher, home-screen banner, and a small Kotlin bridge that forwards remote-control keys to JavaScript so focus is explicit and predictable on a television.

The AI is asked exactly three questions: what did this caregiver describe, what did this person just say, and which approved alternative fits. Every answer is schema-checked with zod. The model never writes to the database — the server maps an understood intent onto a validated command, and replies are composed from published text.

Challenges we ran into

React Native's TV support lives in a separate fork that nothing in the Fire TV materials mentions, so we wrote our own Kotlin key bridge. Sharing a TypeScript package across the monorepo took three build cycles to get Metro to resolve. A cold Gradle build kept hitting HTTP 429 from Maven Central.

The hardest one: our build machine had no hardware acceleration, so no emulator could start. We eventually found that an architecture-matched x86_64 Android TV image boots under pure software emulation in about ten minutes. Running the app there found five real bugs nothing else had caught — including a setup screen that trapped you (dismissing the keyboard left the field holding focus, so the D-pad stopped working) and a help confirmation that never appeared because a helper cleared the message right after the action set it.

What we learned

Running it on a television taught us more in twenty minutes than the tests did in a day. Also that the honest answer is usually the better one: saying "I don't have that written down, and I can't see the room, so I don't want to guess" is a feature, not a limitation.

What's next

Real notification delivery with a tested failure path, escalation when a help request goes unacknowledged, multiple languages, and connecting the MCP server to Alexa+ once Preview access allows it.

What is real and what is simulated

Built With

• android
• express.js
• fire-os
• fire-tv
• kotlin
• model-context-protocol
• node.js
• react
• react-native
• server-sent-events
• sqlite
• typescript
• vite
• vitest
• web-speech-api
• websockets
• zod