How it works

Three engines. One quiet library.

PeetaBoo is not a smarter baby monitor. It's a pipeline that runs on the footage your cameras already produce: it finds your child, writes down what happened, and keeps what's worth keeping.

Input
Camera
Any camera that saves video files, or our own Raspberry Pi camera kit.
01 · Find
Is that my child?
Person detection, face matching, and nine phases of checks. Runs on hardware in your home.
02 · Understand
What happened?
A vision-language model writes plain observations for matched clips only.
03 · Keep
What's worth keeping?
A daily journal on your phone. Bookmark a clip and it never expires.
Engine 01 · Find

Is that my child?

"Is my child in this video?" asked ten thousand times, with a low tolerance for being wrong.

  1. Enrol from 3 to 5 photos. Each face becomes a 512-dimension embedding. Several centroids capture how your child looks frontal, in profile, and in different light, rather than one average.
  2. Find the people first. A person detector we fine-tuned on ceiling-camera footage finds bodies and tells children from adults. Face detection only runs on those crops, so a face on a TV screen or in a photo frame never counts.
  3. Gate on quality. Faces that are too small, blurred, badly lit, or turned away are scored and demoted before they can produce a noisy match.
  4. Match, then check the family. The face is compared against your child, and against a gallery of parents and siblings. If it's closer to a family member, it's rejected, whatever the raw score.
  5. Confirm across frames. One high-quality strict match, or several matches inside a 90-frame window. A single lucky frame is not enough.
  6. Suppress the impostors. Adults, oversized bounding boxes, and things that never move (a doll, a photo) all pull the score down.
  7. Copy, never move. Matched clips are copied into your library and annotated. The original file is never touched.
YOLOv8n, fine-tunedRetinaFaceArcFace 512-dBoT-SORT trackingK-means centroid gallerynegative gallerystrict / review / no match
Engine 02 · Understand

What happened?

"What is my child doing in this video?" answered as an observation, never a verdict.

  1. Only matched clips go further. If Engine 01 didn't find your child, nothing is analysed.
  2. The model knows your household. Its instructions are built from your child's profile: names and roles of the people at home, age in months, anything you've said you're curious about.
  3. Native video, not screenshots. The clip is sent whole, so timing, movement, and sound all count.
  4. Structured output. Each clip becomes timestamped segments: activity, what exactly happened, body position, mood, who was present, how they interacted, and a developmental note grouped by domain (gross motor, fine motor, language, social, cognitive, self-care).
  5. A daily summary. Segments roll up into the day's key moment, activity breakdown, and emotional profile.
  6. Guardrails that don't bend. Low temperature. No clinical labels. Never a diagnosis. Anything notable is framed as something to bring to your pediatrician.
vision-language modeltemperature 0.2JSON segmentssix developmental domainsobservations, not diagnoses
Engine 03 · Keep

What's worth keeping?

A library, not a hard drive.

  1. Every evening, a journal entry. The day's key moment with its clip, an activity breakdown, mood through the day, and a 7-day trend.
  2. Bookmark means forever. Tap the bookmark and the clip is tagged to never expire. Nothing, and nobody, can age it out afterwards.
  3. Everything else expires on a schedule. Retention is enforced by storage rules, not by remembering to delete things. See the table below.
  4. Your household, your rules. Admins run the household. Viewers, like grandparents, see everything or nothing. Invitations are single-use and expire in 48 hours.
  5. More than one child. Each child has a profile, an enrolment, and their own journal.
iOS & Androiddaily journallibrarylive viewadmins & viewersmulti-child
Retention

What happens to every clip.

Each clip carries two tags: the verdict from Engine 01 and whether you've bookmarked it. Storage rules act on those tags. This is the actual policy we run.

ClipVerdict tagKeptNotes
Bookmarked by youkeep = trueForeverSet by your first bookmark. Never reset, even if you remove the bookmark later.
Your child, confidentlystrictForeverMoves to cheaper storage after 30 days. Always playable.
Borderline, needs a lookreview180 daysCheaper storage after 30 days. Promote it, or bookmark it, and it stays.
Not your childno_match30 daysDeleted automatically. A handful are kept as negative examples to improve the detector.
Uploaded, never processedpendingShort graceDeleted if the pipeline never gets to it, so nothing sits around by accident.
Cameras

Works with the camera you already own.

If your camera saves video files we can read, PeetaBoo can process them. We run it on a wall-mounted Reolink in our own living room. For homes without a camera, we're building a small kit of our own.

Bring your own

Any camera that writes files

Reolink, Wyze, Ring, Nest, Arlo, Eufy, or a plain IP camera saving to a NAS. MP4, MOV, and MKV are all fine. No proprietary protocol, no camera upgrade.

  • Motion-triggered clips work best: short, and already filtered to moments where something happened.
  • Wall or ceiling mounts are what the detector was trained on.
PeetaBoo camera kit · in development

A Raspberry Pi that only keeps people

Our own device records motion clips at 1080p30, checks each one for a person on the device itself, and deletes the rest before anything leaves the house.

  • No cloud credentials on the device. It asks the backend for a one-time upload link per clip.
  • Heartbeat every minute: temperature, disk, and anything stuck, so we know before you do.
  • Up to ten days of clips buffered locally if your internet drops.
Raspberry Pi CM5Camera Module 3on-device person check1080p30, 6 Mbps
Privacy

Your child's face never leaves your home.

This started as software for one family's footage. The rules we set for ourselves are the rules the product ships with.

See it on your own footage.

We're in private beta with a handful of families. If you have a camera and a kid under three, we'd like to hear from you.

Oakville, Ontario · a human replies within a day · hello@peetaboo.com