Nightlight

Nightlight learns your household's nights from the Ring doorbell it already has, answers a 3am doorway with a recorded family voice, and wakes the caregiver only when the voice is not enough.

Comment

Inspiration

At 3am, a person living with dementia opens the front door. Most door alarms respond the same way: they wake the caregiver. Night after night.

That caregiver is the second patient in the house, and the night is usually what ends care at home:

  • 6 in 10 people living with dementia will wander at least once [1], and night-time wandering occurs in about 60% of cases [2].
  • Nearly half of dementia caregivers (46.84%) have trouble falling back to sleep after being woken at night [3].
  • 70% of families who moved a relative into a care home cited night-time problems in that decision, often their own lost sleep [4].

Researchers already tested the obvious answer. A night monitoring system that woke the caregiver made people safer, with fewer injuries and unattended exits [5]. But in the companion trial, 49 caregivers were followed for a year and their sleep did not improve on any measure [6]. They felt supported and stayed just as tired, because the system's only response was to wake them.

So we asked a different question: what if the first response to a 3am doorway did not wake anyone?

What it does

Nightlight is software behind the Ring doorbell a family already owns. No new hardware, no wearable, no locked doors.

  1. Learns the household. For its first week it only watches, building a picture of a normal night in this home, hour by hour. It learns new quiet quickly and new activity slowly, so a night of wandering can never teach it that wandering is normal.
  2. Notices the unusual 3am doorway. Only people count. Ring's person detection and doorbell presses can start an alert, so cars, pets and wind never wake anyone.
  3. Answers with a familiar voice first. A message the family recorded plays at the door through Ring: "Dad, it is night time. Come back inside. I will see you in the morning." If no recording exists yet, a natural synthesized voice is used.
  4. Wakes the caregiver only if the voice is not enough. If activity continues, the caregiver gets a notification with a doorway snapshot and a one-tap "I have it". Backup contacts are alerted only if that goes unanswered.
  5. Writes a calm morning note. A short summary of the night, worded by Claude on Amazon Bedrock from facts the engine computed. It never adds facts of its own.

It never fails into silence: if the voice cannot play, the caregiver is woken immediately.

The number on the dashboard is the product: nights the caregiver was not woken.

Does it work? Measured on real homes

We replayed 2,936 nights from 34 real homes in the public CASAS smart-home research dataset [7] through the same engine that runs in production.

Times the caregiver is woken
Standard door alarm 14,068
Nightlight 774

That is a 94.5% reduction, and 365 doorway moments were settled by the recorded voice alone, with nobody woken.

Waking someone less is easy if you ignore everything, so we also measured what that restraint costs. Of 34 night-time exits labeled in the dataset, Nightlight flagged 7 as unusual. Of the 27 it let pass, 26 were exits the resident came back from within thirty minutes. Both numbers, the per-home results and the limits are published in the repository.

Try it yourself

On the live site you can:

  • Send real Ring webhooks to the deployed system and watch it accept a genuine 3am doorbell event, reject the same message with one byte altered, and ignore a duplicate.
  • Ask Alexa how the night went: open a live session from your browser to the household's agent server and see the same answer the caregiver sees.
  • Walk through a demo month night by night. Demo data is simulated and always labeled as such.
  • Get a household of your own with a free API key, no account. Replay a month into it, post your own doorbell's events into it, change its night window and ask an assistant about it, without touching the shared demo. No key still works for everything.

How we built it

  • Ring Partner API: signed webhook intake, device list, event history, snapshots and chime audio, plus the token exchange and account link endpoints Ring certification requires. Verified against Ring's live API and the sandbox doorbell in Ring's Developer Playground.
  • Alexa+ and agents: an MCP server (spec 2025-11-25, Streamable HTTP) with five household tools, so Alexa+ or any agent can ask about the night or acknowledge an alert hands-free. A Strands agent on Amazon Bedrock AgentCore uses it as an independent client.
  • AWS: Lambda, DynamoDB (the event log, with exactly-once actions so the voice can never play twice), Bedrock (the morning note, with a plain-template fallback), Polly (the fallback voice), S3 and CloudFront, all deployed by one CDK stack.
  • Safe by design: the detection engine is rule-based and explainable. No AI model ever decides whether to wake someone.
  • Open source: we extracted the Ring webhook intake into ring-webhook-kit on npm (MIT, zero dependencies), so other Ring developers do not have to rewrite signature checking.

Running cost: under 20 cents per household per month, measured on the live deployment.

Why it matters

A semi-private nursing home room costs $315 a day [8]. A randomized trial that supported family caregivers, without treating the patient at all, delayed nursing home placement by a median of 329 days [9]. Nightlight targets the same thing that trial did: the caregiver's ability to keep going.

We do not claim Nightlight delays placement; that needs a long-term study. What we have measured is a 94.5% cut in night wakings, at a price where one day in that room pays for more than a century of Nightlight.

Challenges we ran into

  • Ring's chime audio endpoint does not document which audio formats it accepts, and that decides how our core feature is delivered. We built a fallback through Echo announcements and reported the gap to Ring.
  • An early version could slowly learn a wandering night as normal. We redesigned learning to be one-sided so that can never happen.
  • Two of the Bedrock models behind our morning note reached end of life and failed silently, hidden by a first choice that still worked. Every fallback is now checked, not just the first.

Accomplishments that we're proud of

  • A headline number that counts wakes that did not happen, measured on real homes and published next to what it missed.
  • A real Ring integration, not a mockup: live API calls, signed webhooks and a working sandbox device.
  • One deployment serving both the Ring and the Alexa+ experience from the same household data.

What we learned

Restraint is a feature. The hard part was never detecting a door opening. It was earning the trust to stay quiet.

What's next

Ring certification into the Appstore's elderly care category, support for homes with more than one door, and a caregiver sleep study with a dementia care organization.

Nightlight is a home safety aid. It does not diagnose or treat any condition and does not replace supervision or professional care.

Sources

  1. Alzheimer's Association, Wandering
  2. Serban et al., Alzheimer's & Dementia, 2025, PMC12725205
  3. Osakwe et al., Frontiers in Aging Neuroscience, 2022, full text
  4. Pollak and Perlick, Journal of Geriatric Psychiatry and Neurology, 1991, DOI
  5. Rowe et al., Alzheimer's & Dementia, 2009, PubMed
  6. Rowe et al., Journal of Nursing Scholarship, 2010, abstract
  7. CASAS HH dataset, Washington State University, Zenodo 15708568 (CC BY 4.0)
  8. CareScout Cost of Care Survey, 2025, results
  9. Mittelman et al., JAMA, 1996, PubMed

Every claim with its full citation: EVIDENCE.md

Built With

Share this project:

Updates

Submission history