Whisker

Full Hardware Setup on Baby Crib
Sensors Top View
Hardware Placement Parts
Hardware Placement Parts
Coral TPU Connected to Raspberry Pi
Raspberry Pi

Inspiration

Babies in the NICU are covered in wires. Every lead, probe and cuff is there to track a handful of vitals: breathing, temperature, weight and movement. We realized most of that can be measured from a distance with sensors built into the crib, without touching the baby at all. Whisker brings that kind of monitoring home, so parents can sleep knowing someone is still watching.

What it does

Whisker is a smart crib that monitors a sleeping baby and alerts parents when something goes wrong.

Tracks breaths per minute, baby body temperature, room temperature, weight and movement, all without anything attached to the baby.
Streams a live night vision video feed to a web dashboard.
Shows live vitals cards, historical charts (up to 30 days), a night timeline of activity and alerts, and a full alert log with a snapshot of the crib at the moment each alert fired.
When a vital crosses a danger threshold (stopped breathing, unsafe body temperature, unsafe room temperature, or the baby leaving the crib), the dashboard shows a full screen alert and Whisker places an automated phone call to the parent.
Everything sits behind login, so only the family can see the feed and the data.

How we built it

Hardware and edge (C++ on a Raspberry Pi)

A Coral Edge TPU runs an SSD MobileNet v2 detection model on the camera feed to find the baby in the crib and track movement.
The detected position drives a Pololu Maestro servo mount that aims a TF-Luna lidar at the baby's chest. We filter the distance signal and count the rise and fall of the chest to get breaths per minute.
An MLX90640 thermal camera is sampled at the detected location for body temperature, and the room temperature comes from the same sensor.
Pressure sensors under the mattress, read through an ADS1115 ADC, give weight and tell us whether the baby is in the crib.
The Pi writes readings to a TimescaleDB hypertable on Tiger Data, and sends video frames and emergency packets over an authenticated TCP connection to our server.
All enclosures and mounts were designed and 3D printed by the team.

Backend (Flask on AWS EC2)

Pulls the latest readings and time bucketed history from Tiger Data with plain SQL and time_bucket().
Pushes live vitals to the dashboard over Socket.IO and relays the video feed.
On an emergency packet it saves the alert and snapshot to the database, pushes it to every open dashboard, and calls the parent through Twilio Voice at the same time.
Served by gunicorn behind nginx with HTTPS on our own domain.

Frontend (React + Vite)

Live vitals cards, Recharts history charts, a draggable night timeline with alert pins, a 3D animated baby that reflects movement, and a full screen alert modal.
Auth0 login with JWTs checked on every API call and socket connection.

Challenges we ran into

Aiming a single point sensor at a moving baby. The lidar only measures one spot, so the detection model, servo control and breath counting all had to stay in sync in real time.
Noisy signals. Breathing shows up as a few millimeters of chest movement. We had to smooth the signal and reject noise without missing real breaths.
Getting video into a browser. The Pi sends raw frames over TCP, and browsers cannot read that directly, so the backend had to relay it.
False alarms. A sensor reading of zero can look like a baby that stopped breathing. We had to separate "no data" from "danger" so parents are not woken up for nothing.
Twilio trial limits. Verified numbers only, blocked inline call scripts and region permissions all broke our first calls on the way to a working setup.
Building in parallel. We built the dashboard against simulated data with the exact same shape as the real database, so switching to live sensor data was a one line change.

Accomplishments that we're proud of

A full pipeline from sensor to phone call: crib, Pi, cloud database, server, dashboard and voice call, all working live.
A working monitoring camera that detects and follows the baby's movements precisely.
Measuring breathing with no contact with the baby, using a lidar aimed by a detection model running on the edge.
An alert reaches the dashboard and starts a phone call within seconds of the crib detecting it.
A polished dashboard that a tired parent can read at a glance at any time necessary
Custom 3D printed hardware built from scratch in 24 hours.

What we learned

Running effective object detection on a Coral Edge TPU and using its output to steer physical hardware.
Turning noisy sensor signals into numbers people can trust, and why alert logic needs to tell missing data apart from real danger.
Storing and querying time series data with TimescaleDB hypertables and time_bucket().
Real time web apps with Socket.IO, and deploying a full stack on EC2 with nginx, gunicorn and HTTPS.
Securing every endpoint and socket with Auth0, and automating phone call alerts with Twilio.

What's next for Whisker

A mobile app with push notifications alongside the phone calls.
Escalation to a second caregiver if the first alert is not acknowledged within five minutes.
Validating our breathing and temperature readings against medical grade monitors.
Sleep insights for parents, such as nightly sleep quality and trends over weeks.
Support for multiple cribs and caregivers in one account.

Built With

Submitted to

CodeRED: Orion
- Winner TVs | Best Hardware Track

Created by

I built the frontend for Whisker. I set up the React + Vite app from scratch, including login, the main dashboard, settings, and the analytics pages for breathing, temperature, weight and activity. I mocked up and applied styles sitewide, including custom logos, page transitions, and loading animations. I also fixed the chart time axes so they show clean, readable times with AM/PM markers. I also built the alert log with a way to clear it (with a confirm step so nobody wipes it by accident), and cleaned up the settings and alert screens near the end.
I wrote the API client and the hooks that stream live readings from the crib, so the dashboard updates as the sensors send data. I also wrote tests along the way too, including contract tests that check the frontend and backend agree on the data format.

Spencer Ennadi-Johnson
I build the edge device for Whisker.
First, I electrically diagrammed and integrated the thermal sensor, IR camera, Coral TPU, single-point lidar, Pololu Maestro, two 25-kilo servos, Raspberrypi 4B, and the pressure sensors.
Then, I programmed all of the drivers to integrate these devices into the C++ edge codebase to then use to track a baby as it moved throughout the crib and to then collect important diagnostics to store into Tiger Data.
Finally, I 3D-modeled all of the parts to for the final device in Fusion360 and then printed the parts with a Flashforge 5M.

Matthew Beck
Robotics & Embodied AI, ML/Deep Learning, Data Engineering & Backends, Cloud (AWS)
I worked on the backend data pipeline and the Twilio integration. Before the sensors were live, I built a dummy data generator that produced realistic readings for breathing rate, baby temperature, room temperature, and weight, in the exact format the real Tiger Data table would use. This let the frontend be built and tested in parallel with the hardware, and switching to live data was a one-line change. Using Flask, I built the pipeline that polls our Tiger Data (TimescaleDB) table every 30 seconds for the latest vitals and pushes only new readings to the dashboard in real time over Socket.IO. Since the board writes about once a minute, each reading reaches parents exactly once, within 30 seconds. I also built the history endpoints for the charts, which use TimescaleDB’s time_bucket() to average readings over ranges up to 30 days. I built the Twilio voice-call integration for emergencies. When an alert comes in, the backend calls the parent’s phone with an automated message explaining the specific problem, such as a high body temperature or a very low breathing rate. I also worked with Allen to connect the alerts from the board to the dashboard, so they appear as soon as the board detects them. On the frontend, I used Three.js to create a 3D baby model that moves when the crib detects real movement and lies still when there is less movement.

Vishnu Muthyalu
I build the backend pipeline and system admin for Whisker

Built the end-to-end alert pipeline, from C++ TCP communication on Raspberry Pi to Flask backend processing. Implemented three concurrent alert-handling paths for emergency call dispatch, real-time Socket.IO notifications, and TigerData persistence. Also contributed to Flask/Socket.IO backend development and managed cloud infrastructure, including domain setup, HTTPS certificates via Certbot, Auth0 API configuration, and JWT token handling.

Allen Wu