Inspiration
It's 3 AM. A baby won't stop crying, the mother is exhausted, and there's no one to ask. I'm an aunt of two little ones, and I've watched my sister and my friends go through those nights. In Pakistan and many other countries, help is often far away, most health apps are in English, and well-meant family advice isn't always safe: studies in Pakistan have found many families discard the first milk (colostrum) or give newborns ghutti or honey.
I wanted to build the village every new mother deserves: someone kind to ask, any time, in her own language.
What it does
Motherhood is a voice-first AI companion for pregnancy and the newborn months.
- Talk or type, any time. She taps "Talk to me" and speaks; answers stream in within seconds and are read aloud, in short, warm sentences.
- Knows her and her baby. A 30-second setup (no password) makes every answer right for the baby's age or the pregnancy week. A "Today" card shows what's normal right now.
- Voice baby log. She says "she fed 20 minutes on the left and two wet diapers" and Nemotron's tool calling logs it. The app gently warns if diapers or feeds look low for the baby's age.
- Safety first. A red Emergency button shows danger signs and one-tap calls to local numbers (Rescue 1122 and Edhi 115 in Pakistan). Crisis words in five languages open it automatically, even if the AI is down.
- Weekly mood check-in, based on the Edinburgh Postnatal Depression Scale, to catch postpartum depression early.
- 3D picture guides for latching, burping, swaddling and safe sleep.
- Myth or safe? Respectful answers about family traditions like ghutti, honey and surma, with a "Share with family" button.
- Voices from your village. She saves voice notes from people she loves, even a mother she has lost, and hears their real voice when she says she's tired or sad.
- A library of 44 gentle articles, plus pregnancy week by week.
- Five languages: English, Urdu, Arabic, Spanish and Turkish, with right-to-left layouts.
- Private by design: her profile, log, check-ins and recordings stay on her own phone.
How we built it
- NVIDIA Nemotron 3 Nano (
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B) writes every answer, running on Nebius Token Factory through its OpenAI-compatible API. - Streaming for fast first words, with reasoning switched off for speed and quiet automatic retries, so a mother never sees an error.
- Tool calling (
log_baby_events) turns spoken sentences in any of the five languages into structured baby-log records. - Personal context: the baby's age, feeding type and last 24 hours of the log are added to each request.
- On-demand translation of articles, guides and cards by Nemotron, with checks that reject wrong scripts (for example Hindi letters in Urdu).
- A zero-dependency Node.js server; a single-page progressive web app with the Web Speech API, IndexedDB, MediaRecorder and a service worker; Three.js for the 3D guides.
- A server-side crisis safety net that works before, and without, the AI.
Challenges we ran into
- Urdu done right. The smaller model sometimes slipped into Hindi script, so we added script checks and retries. Free text-to-speech models read Urdu with a Hindi voice, so we chose to use only real Urdu voices rather than get it wrong.
- Keeping answers short. Tired mothers need three to five sentences, not essays, so we combined prompt rules with a server-side limit that ends cleanly on a full sentence.
- Medical accuracy. Every safety fact (formula preparation, milk storage, fever under 3 months, danger signs) was checked against NHS, CDC, WHO and AAP guidance, and the latch guide was redrawn until the positions were right.
- Access from Pakistan. Some voice services needed phone verification or payment methods not available here, so we designed features to work well without them and fall back gracefully.
Accomplishments that we're proud of
- A safety net that never depends on the AI being up.
- Respectful handling of family traditions, so elders feel honored, not corrected.
- Five languages, including proper right-to-left Urdu and Arabic.
- Voices from your village: a feature built around love, not just information.
What we learned
Small, fast models can be wonderful companions when they're given clear rules, real context about the user, and guardrails around them. And the hardest problems weren't technical: they were about trust, language, and getting medical details exactly right.
What's next for Motherhood
- Testing with mothers, Lady Health Workers and doctors in Lahore.
- Hearing guidance in a loved one's voice, with consent built in (we've applied for NVIDIA Riva zero-shot voice access).
- More languages, including Punjabi and Pashto
- A health-worker mode to help community health workers check danger signs during home visits.
Built With
- chatterbox
- css3
- github
- html5
- hugging-face
- indexeddb
- javascript
- mediarecorder
- nebius-token-factory
- node.js
- nvidia-nemotron
- openai-api
- pwa
- render
- service-worker
- three.js
- web-speech-api
Log in or sign up for Devpost to join the conversation.