Privacy-First Intelligence

Your phone learns
your patterns

Pre-trained on peer-reviewed behavioral research. Personalized on your device. No data ever leaves your phone.

The problem

Social media is designed
to keep you scrolling

42%

of people say they can't spend a week without social media.

Social Media Behavior Survey, 2024. N=310.
76%

get distracted by social media while working or studying.

Social Media Behavior Survey, 2024. N=310.
84%

believe platforms are actively monetizing their attention.

Social Media Behavior Survey, 2024. N=310.
52%

don't use any tool to manage their digital distractions.

Social Media Behavior Survey, 2024. N=310.
Our approach

Trained on research.
Personalized to you.

Stilla's prediction engine starts with published behavioral science and adapts to your unique patterns over time.

Research Foundation

Pre-trained on 26,710 real-time desire episodes from the Qwantify ESM study (636 participants). Peer-reviewed, open data.

Key Discovery

Time of day barely matters. What predicts cravings: rumination (29%), stress (14%), hunger (10%), and bad mood (7%). Your mental state drives the urge.

On-Device Learning

As you log cravings, the model adapts. After 10 entries, it starts personalizing. After 50, your own data dominates. The research priors fade as your patterns emerge.

28 Features

The model considers time, activity, mood, energy, stress, loneliness, rumination, hunger, physical state, and 19 emotional indicators — all from your own logs.

Smart Notifications

Pre-emptive alerts before your peak craving windows. Context-aware messages that adapt to your behavior. Stays quiet when you're doing well.

Zero Data Leaves

Everything runs on your phone. No cloud. No server. No analytics. Your craving patterns, predictions, and model are yours alone.

How it works

Three phases of
personalization

Day 0

Research priors

Predictions based on published behavioral science. Generic but directionally accurate. 15% confidence.

Day 1–30

Blended learning

Your craving logs gradually override the research data. The model learns YOUR triggers, YOUR peak times, YOUR best interventions.

Day 30+

Fully personal

Your data dominates. The model knows your patterns better than you do. Predictions are tailored to your life.

The science

What actually
predicts cravings

From 26,710 real-time desire episodes. Not surveys — actual moments logged as they happened.

29%

Rumination — overthinking is the strongest predictor of desire.

Qwantify ESM dataset. Gradient Boosting, 68.7% accuracy.
14%

Stress — the second strongest driver of craving behavior.

Qwantify ESM dataset. Feature importance analysis.
10%

Hunger — being hungry significantly increases craving probability.

Qwantify ESM dataset. Physical state features.
7%

Bad mood — low mood predicts desire, but less than overthinking.

Qwantify ESM dataset. Mood scale (0–1).
Open science

Built on open data

Our model is trained on the Qwantify experience sampling dataset — 636 participants, 26,710 desire episodes, published under open access.

Hasenkamp, W., Wilson-Mendenhall, C., Condon, P. (2022). Frontiers in Psychology.
Dataset: osf.io/sxfrx