Align

Inspiration

Every startup and creator is building agents that connect to our devices, our apps and our accounts. But no agent connects to what matters most: our personal life — our family, our health, our relationships. Life drifts quietly: work eats the evenings, the gym goal dies in week two, date night disappears, and nobody notices. We built the agent that does.

What it does

Align has one goal: to make you as successful as you can be, in every part of your life — family, social life, health, work and personal growth.

It keeps a living memory of your life — your people, goals, preferences and commitments — each with a confidence score. That score updates with every conversation and every signal Alexa shares with it, so Align always knows what it's sure of and what it still needs to learn.

From that memory, Align:

• Speaks first when something matters — and stays quiet when nothing does.
• Asks one smart question at a time, only when the answer would change the plan, moving across every area of your life.
• Builds one plan for your whole week, respecting what it learned about you ("I hate morning gym").
• Acts only with your yes, and verifies every change.
• Measures what really happened — a plan on the calendar isn't a plan done — and learns from it.

How we built it

• A real MCP server for Alexa+: 13 tools, Streamable HTTP, spec 2025-11-25.
• A deterministic engine makes every decision (8 pattern detectors, question scoring, planning, verification); the language model (Claude) only speaks.
• A simulated Alexa+ host, clearly labeled, because agents that speak first aren't open to developers yet.
• 146 automated tests, including an evaluation suite for the intelligence itself.

Challenges we ran into

Alexa+ has no API yet for agents that speak first, so we built and labeled our own host channel. Our new AWS account was blocked from Amazon Bedrock capacity (AWS Support: new accounts stay limited), so we voiced the demo through the Claude API — without changing a single decision in the engine.

Accomplishments that we're proud of

An agent that provably learns.

What we learned

• An LLM alone can't run a life. Our first versions let the model decide, and it invented changes, repeated itself and forgot what the user said. Moving every decision into a deterministic engine, with the model only speaking, made Align reliable, fast and auditable.
• Voice changes everything. Long answers kill a voice assistant. We cut every reply to two sentences, spoke times the way people say them ("7 pm", not "19:00"), and made Align ask one question at a time.
• A plan on the calendar is not a plan done. Counting scheduled sessions as success made the score lie. Only confirmed evidence should move it, so Align learned to ask "how many really happened?" instead of guessing.
• Memory has to be honest. Every fact needs a confidence score, a source, and a way for the user to correct or delete it. Otherwise trust breaks the first time the agent is wrong.
• Build so no single provider can stop you. When our new AWS account was blocked from Bedrock, the MCP architecture let us swap the voice to the Claude API in hours, without touching the engine. ```

What's next for Align

Align today is a working prototype of a simple idea: an assistant that knows your whole life and helps you live it better. Our next step is to build it with the Amazon Alexa team, turning this prototype into a production-grade architecture: a deeper, long-term memory, smarter reasoning across every part of life, and native agent-initiated conversations on real Alexa+ devices.

The vision is "Alexa, by Align": first an assistant that manages the entire life of one person (calendar, health, goals, relationships, growth) with their consent, then the whole family, with Voice ID giving each member their own memory, goals and privacy, and plans that bring the household together. One goal, for everyone at home: a life at ten out of ten.

Built With

• amazon-bedrock
• amazon-ecr
• amazon-ecs
• aws-fargate
• claude
• docker
• model-context-protocol
• node.js
• sqlite
• typescript
• vitest
• web-speech-api
• zod