# MaPa

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Shreyash Sharma](https://cerebralvalley.ai/u/Shreyashs)
- **GitHub:** https://github.com/shreyash-sharma/MaPa
- **Demo video:** https://www.youtube.com/watch?v=dO7K1QtBNBU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3129

MaPa - A Little Closer, From Anywhere

What if you could know when your parent's day looks different, without constantly calling or monitoring them?

I live in Pune, while my parents live in Bhopal. Like millions of people living away from their ageing parents, I wanted a way to stay connected without compromising their independence.

So I built MaPa, a privacy-first Android application that learns each parent's individual phone-usage rhythm and identifies unusual changes.

Unlike fixed inactivity timers, MaPa understands that everyone's routine is different. My father wakes up early; my mother's mornings look completely different. Neither should be judged by the same arbitrary threshold.

 What makes MaPa different?

1. Original, lightweight AI

We trained a 9,594-parameter neural network from scratch, combining a causal CNN, GRU, 30-day LSTM, and personalized profiles. Using Meta's ExecuTorch and XNNPACK, we developed an on-device inference path.

2. Our own research dataset

Existing datasets weren't sufficient for our target problem. We combined public smartphone research data with 200 statistically modelled parent timelines and ten types of controlled behavioural scenarios, separating training and evaluation by person.

3. Meta-powered multilingual intelligence

Meta's Muse Spark Open, served through Baseten, powers daily summaries, questions, language assistance, and life-context tags. The language model explains information but never controls the anomaly-detection decision.

4. Privacy by design

No cameras, wearables, GPS, messages, or app names. Raw phone observations remain local. Parents explicitly consent and can pause sharing. Missing data is never presented as evidence that everything is fine.

5. Measured and reproducible

On 40 new model-derived parent timelines, our latest decision policy raised alerts in all 108 labelled behaviour windows, with 1.09 false alerts per observed person-week versus 6.54 for a fixed four-hour timer.

These are controlled simulation results, not proof of real-world emergency detection. We also document unresolved false-alert problems.

 Built openly

We've released the Android application, trained model, dataset, ExecuTorch exports, training pipeline, and evaluation artifacts through GitHub and Hugging Face.

MaPa isn't intended to tell families whether someone is safe. It's intended to give them a better reason to check in.

Because our parents deserve both their independence and our care.

## More from Meta Global AI Developer Hackathon

- [SCMTIPS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3126)
- [HalalNumeriQ](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3127)
- [JeevShastra](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3128)
- [Team Promptimists](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3130)
- [EveryMuse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3131)
- [Inkling](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3132)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3129. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt