Bio Engine — the four scores
The Bio Engine combines signals from your phone's sensors, your wearable (if you have one), and your meal, sleep, and activity logs into four scores. Each is a 0–100 indicator computed locally on your device.
Neural Battery
What it is: A 0–100 estimate of mental energy available right now.
Inputs: Last night's sleep duration and quality, HRV (if available), screen time patterns over the last 6 hours, recent mood logs, time since last meal.
What it's not: A clinical measure. It's a directional indicator based on the best consumer-grade signals we can read. Use it to ask "should I push hard today or rest?" — not to diagnose anything.
Hormonal Load
What it is: A 0–100 estimate of stress-system activation, derived from a cortisol proxy.
Inputs: Sleep onset latency, resting heart rate vs. your baseline, heart rate variability, recent workout intensity, and self-reported mood.
What it's not: A direct cortisol measurement. A proxy built from validated consumer signals. The trend matters more than the absolute number.
Physical Fatigue
What it is: A 0–100 estimate of accumulated physical strain.
Inputs: 7-day training load (TRIMP-style), sleep debt, deep sleep share, recent workout intensity, RPE (if you've logged it).
Vitamin Analysis
What it is: A daily estimate of 8 key micronutrients from your meal log.
Inputs: Foods logged (camera or text), serving sizes, our nutrient database (over 800,000 foods).
What it's not: A medical-grade deficiency diagnosis. It's an estimate from the food you tell us about — what you don't log, we can't see.
Important: Body Mode is a wellness companion. None of these scores are medical-grade. They are designed to help you make better daily decisions, not to diagnose, treat, cure, or prevent any disease. Always consult a healthcare professional.
Sleep without a wearable
The phone-based sleep detector fuses four signals into a draft sleep event every night. The event is presented to you in the morning for one-tap confirmation. Confirmed events write to Health Connect.
- Accelerometer: motion stillness over rolling 10-minute windows
- Screen state: screen-off during your learned sleep window is a strong signal
- Charging state: charging overnight correlates strongly with sleep periods
- Time-of-day: your learned sleep window (updated each night you confirm)
Validation: We've tested this against wearable-grade sleep trackers in 200+ nights of mixed-user data. Mean absolute error on total sleep time is under 30 minutes, on par with consumer wrist wearables.
Adaptive plan generation
Every morning, the plan generator reads your last 7 days of data and builds a single day plan. The plan is generated by a large language model grounded in your data, with strict post-generation constraints applied to ensure availability windows, learned meal counts, and intensity preferences are respected.
Why a language model: Plans are not just numbers. They need to reason about your life context — "you have a meeting at 2pm so lunch should be lighter", "you had a hard workout yesterday so dinner should emphasize protein and recovery", "you said you're traveling so I'll suggest a hotel-room workout instead of a gym session." A rule engine can't do this well. A language model can.
What grounds the model: Your last 7 days of meals, sleep, workouts, mood, and goals. Your stated preferences. Your current health metrics. The model sees all of it; you only see the result.
Privacy by architecture
Most of what Body Mode does — sleep detection, Bio Engine scoring, food logging, plan generation context — happens on your device. Only the final plan generation step requires a model call, and the data sent is the minimum context needed to generate your plan. No raw sensor data leaves your phone.