Inspiration
We got inspired by a real problem. Wilma stayed up the whole night when the neighbour's baby was crying nonstop. It made us wonder: if a baby cries like that, how do the parents know what's wrong?
So we asked. We interviewed 8 parents, mostly in Finland, via WhatsApp. 7 out of 8 said they struggle with unexplained crying and end up relying on trial and error. 6 out of 8 struggle with sleep and exhaustion. And every time someone new takes care of the baby, whether that's the other parent, a grandparent or daycare, the guessing starts from zero. One parent put it perfectly: "What worked for other parents didn't work for me." Another said she doesn't want any more time on her phone.
What it does
Waby is a wristband for babies that listens, plus a shared page every caregiver can open.
- Hears crying. A sound model recognises a baby's cry within a couple of seconds.
- Suggests why. "Probably hungry: last fed 2 hours 50 minutes ago, usually eats every 3 hours", with what to try first, based on what calmed this baby before.
- Logs care by voice. A caregiver says "I gave her a bottle, four ounces" near the band and it is in the log. No typing, no extra screen time.
- Answers out loud. "When did she last eat?" gets a spoken answer in about 4 seconds.
- Hands over. A new caregiver asks "What did I miss?" and hears a short summary of the day.
- One shared timeline. Anyone with the link sees the day's key moments. Nothing to install.
Waby does not claim to read meaning from the cry itself. The likely reason comes from this baby's own history: time since the last feed, time awake, the last diaper, and what ended earlier cries.
How we built it
- Band: the Vexo Open Kit wristband. Its microphone streams Opus audio over Bluetooth LE.
- Gateway: a Python client on a laptop, built on the Vexo open kit library. It stands in for a phone or home hub.
- Cry detection: YAMNet, Google's sound classifier trained on AudioSet, running locally on the gateway with LiteRT. No audio leaves the laptop for this step. Tested through the band in the hackathon room: it caught a real cry recording played at 40% phone volume and gave zero false alarms over 3.5 minutes of people talking.
- Speech: Deepgram Nova-3 for streaming speech to text, Deepgram Aura-2 for the spoken answers.
- Backend: one Cloudflare Worker with a D1 (SQLite) database that holds the care log.
- Understanding: Llama 3.3 70B on Cloudflare Workers AI with structured output. One call decides whether speech is a note to log, a question to answer, or chatter to ignore. The likely reason for crying starts from plain rules, so it is instant and still works if the model is slow.
- Interface: a static web page served by the same Worker, with a live transcript, spoken answers and a timeline of key moments.
Challenges we ran into
- A loud room. Speech to text never heard a pause, so a question was never "finished" and never sent. We now send a question the moment it ends.
- Speed. A spoken answer took up to 10 seconds. We measured each step and got it to about 4.
- Phones block sound that starts without a tap. One tap now unlocks an audio engine that stays allowed.
- Telling a question from a note. "When did she last eat?" was first saved as a note. Questions are now answered, never logged.
- The model misread the log. It said "50 minutes ago" for 2 hours 50 minutes. We now hand it the durations already worked out.
- Separating crying from adult speech and background noise. Loudness alone failed, which is why we moved to a sound classifier. It only recognises real baby cries, not adults imitating one, so the demo uses a real recording.
- Bluetooth. The band takes one connection at a time and resets on disconnect, so the gateway reconnects and restores its streams by itself.
- Resisting claims we can't back up, like "your baby is hungry." Answers say "probably" and show the evidence.
Accomplishments that we're proud of
Going from a real problem to parent interviews to a working prototype in one day. Also we think that this is not a typical use case.
What we learned
Parents don't want more data or more screen time. They want to understand their own baby. Every baby is different, so the value is in each baby's own history, not generic advice.
What's next for WABY
- Testing cry detection on real babies and measuring its accuracy.
- Learning each baby's recurring patterns once there is enough reliable data.
- Using the band's motion sensor. We prototyped a restlessness signal and left it out because we could not validate it on a baby.
- Voice from the wrist once the band has a speaker. Today the phone speaks.
- Replacing the laptop gateway with a phone integration or small home hub.
- Validating infant wearability and safety before any real-world use.
- Later: sleep insights and other health signals, each validated separately.
Built With
- bluetooth
- cloudflare-d1
- cloudflare-workers
- deepgram
- javascript
- llama
- python
- vexo
- web-audio-api
- yamnet