Safe-8-Sound

A home battery that keeps the people you love safe

Comment

Inspiration

Heat kills nearly half a million people a year (WHO). A multi-day blackout during a heat wave more than doubles that risk: in a city like Phoenix, over half the population would need medical care (Stone et al. 2023, Environ. Sci. Technol., DOI 10.1021/acs.est.2c09588). Home batteries are spreading fast. In Italy alone, ANIE/Terna GAUDÌ counted 941,911 home storage systems as of 30 June 2026. Yet battery apps only tell you how many hours the battery lasts. Nearly half of indoor heat deaths are discovered during a welfare check (48%, Maricopa County 2025 Heat Report). So Safe-8-Sound tells you how many hours the person in the safe room stays safe, and warns the family while there is still time to act.

What it does

Safe-8-Sound watches one room, the "safe room", for the person who needs it most. Wherever you live, it reads the local weather and risks and builds the safest plan for that home.

It is not only for emergencies. On ordinary days it paces the home battery so the safe room stays comfortable for as long as possible. When the panels make more than the home needs, it keeps that solar surplus at home: it cools the rooms that need it most and suggests when to run the washing machine or do the cooking, instead of sending the energy to the grid.

On a hot day it pre-cools the safe room with surplus solar power and leaves a storm reserve untouched for later. A deterministic guard checks every plan before you see it, so you only see a plan that holds. Options the guard turns down on its own go to the log; they are never shown or spoken. If you ask for something that would eat into the storm reserve, Alexa says so, gives the reason and the numbers, and nothing changes without you.

If a storm knocks out the grid, it steps the home down one comfort level at a time and tells you until what time the safe room stays within safe limits. The family gets a message when a storm is expected, when it starts, when the battery reaches its essential level and when the power is back. If the battery runs low, it offers a last-resort plan that needs two separate confirmations, and it alerts the family at the same moment. It remembers preferences from one day to the next. The AI proposes; a deterministic guard decides.

The family stays in charge through the settings. They choose who is at home, a frail person (in the demo, Leo, an older man) or no one at risk, and which room is the safe room; names, plans and messages follow that choice. Under "Temperatures and limits" they set the comfort temperature and pick °C or °F. The night health limits (26 °C, or 24 °C with an older person) are shown with their sources, and they can only be made stricter, never looser.

Why a storm reserve

A thunderstorm does not always mean a blackout. The storm reserve is a precaution, and how much it matters depends on where the home is.

Piedmont, Italy (our reference home is near Turin): outages are rare and short. In 2023 the main distributor in Piedmont averaged about 21 minutes without power per customer, over 1.19 interruptions (ARERA; low voltage, towns over 50,000 people, exceptional events excluded). Heat strains the grid itself, though: in hot spells underground cables and joints fail as air conditioners push the load up (Il Post, 22 June 2026). And heat is the bigger danger: the summer of 2022 caused 18,010 heat deaths in Italy, the highest rate in Europe (Ballester et al., Nature Medicine 2023). Here the reserve costs little to keep and covers the rare night when the power does go. Houston, Texas: between 2000 and 2023 Texas had more major weather-related outages than any other US state (210; Climate Central, 2024). Hurricane Beryl (8 July 2024) left more than 2 million CenterPoint customers without power, and 88,000 were still in the dark eight days later (Texas Tribune). At least 36 people died in Texas, at least 13 of them from heat or loss of power (ABC13/AP, 26 July 2024). Here the reserve is what keeps the safe room cool through a long, hot outage. Sources: https://www.arera.it/fileadmin/allegati/dati/ele/2023_ambiti_territoriali_alta_tab1_01.pdf · https://www.ilpost.it/2026/06/22/rete-elettrica-perche-ci-sono-blackout-caldo/ · https://www.nature.com/articles/s41591-023-02419-z · https://www.climatecentral.org/climate-matters/weather-related-power-outages-rising · https://www.texastribune.org/2024/07/08/hurricane-beryl-texas-damage-updates-rain/ · https://abc13.com/15096054

How we built it

An MCP server (spec 2025-11-25, Streamable HTTP, stateless on AWS Lambda) that exposes the planning tools, plus MCP Apps cards for the UI. A Strands agent on Amazon Bedrock AgentCore, running on Amazon Bedrock Nova 2 Lite. Spoken replies on the judges' demo site use Amazon Polly (neural voice, behind a cached and capped adapter with a kill switch), which also voices the demo video. If Polly is unavailable or hits its daily cap, the site falls back to the browser's own speech synthesis. Amazon DynamoDB for state; AWS CDK for all the infrastructure. A deterministic guard, not the LLM, owns every safety and comfort guarantee. It is covered by property tests and mutation testing. A simulated Alexa+ experience on top (our self-hosted MCP server behind a simulated Alexa+ voice UI), plus a bridge to a classic Alexa Skill, "safe eight sound", that answers directly in the Alexa Developer Console.

Claims and evidence

The plan keeps every guarantee it promises: guard property tests and mutation testing (≥80%). The plan measurably reduces risk, not only discomfort: docs/results/storm-nights.md covers 11 real nights (the 5 hottest in Turin, the 5 hottest in Houston, plus the Hurricane Beryl night), plan against no plan, with the same weather and physics for both. The numbers Alexa speaks match the numbers on screen: one shared formatter. The storm reserve is never spent behind your back: the guard turns down any plan that would dip below it, and Alexa tells you when your own request would. Only a plan that holds is shown: plans the guard turns down stay in the log and are never shown or spoken, unless the person asked for them.

Honesty box

The hot-day and storm scenarios run on simulated weather and a simulated grid outage, with the real parameters of a home solar and battery system; judges see only the simulator. Each of the 11 storm nights in our study (docs/results/storm-nights.md) is tested on its own from a given starting charge (60%, 35% or 20% at 22:00); solar recharge during the day between nights is not modelled. With one adult in the safe room, the plan keeps 11 of 11 nights within the night threshold, against 9 of 11 with no plan. With one older person (a lower, stricter threshold), neither line keeps a whole night within it, because the room is already above that threshold at nightfall. The plan still cuts the hours spent above it by roughly 70%, and reaches the guard's own recovery target (at least 6 of 9 hours within limits) on 10 of 11 nights at 20% battery, against 0 of 11 with no plan. The storm scenario uses the public outage figures of Hurricane Beryl (Houston, 8–9 July 2024), which left more than 2 million CenterPoint customers without power (Wikipedia, "Effects of Hurricane Beryl in Texas"; S&P Global, 8 July 2024).

If the live agent is unavailable, the site switches to a scripted backup, and every figure from it is marked "Offline example", so nothing is presented as live when it is not.

Here Safe-8-Sound runs on a simulated home, so judges can try every scenario safely. It is built to run on a real one too: the same MCP server will sit next to Home Assistant on a small home server, reading the real inverter, batteries and air conditioners. That home setup is not part of this submission. OAuth 2.1 is the planned next step for the real Alexa+ toolkit, which is open to select partners only and not yet to us.

Challenges we ran into

The Alexa+ MCP Toolkit is "available to select partners only, US only". We never got access, so we built a simulated Alexa+ experience on our own self-hosted MCP server. The MCP TypeScript SDK validates structuredContent against outputSchema even on error results, contrary to its own code comment. The fix was one line; finding it took much longer. AWS promotional hackathon credits require the Paid plan, and the credit request form does not say so.

What we learned and what's next

We learned to keep "what the AI proposes" apart from "what is guaranteed". A deterministic guard, covered by property and mutation testing, is what lets us promise hours of safety and not only hours of battery. Next: when no one at home is at risk, choosing several rooms to keep cool in the settings, not only one safe room; a real Alexa+ add-on with OAuth 2.1, live home control with consent, a community version, and a winter mode in which the same guard keeps the safe room warm during a cold-wave blackout (not built for this submission; summer stays the focus here).

Built With

Share this project:

Updates

Submission history