Privacy-first multimodal fatigue intelligence

Understand today.
Recover better.
Perform tomorrow.

Somno turns reaction time, on-device facial and ocular signals, subjective sleepiness and sleep history into a personal view of fatigue, recovery and readiness.

4signal families
100%on-device face processing
Personalbaseline inference
Somno recovery screen
Somno daily state screen
Reaction speed Personal baseline
State estimate SDI + confidence
React NativeTypeScriptML KitKotlinSupabaseLocal-first

How Somno works

Four signals. One personal state.

Somno treats fatigue as a changing within-person state instead of reducing sleep to a single duration metric.

01

Psychomotor vigilance

A short reaction-time probe captures response speed, lapses, false starts, variability and time-on-task behavior.

02

Face and ocular scan

On-device computer vision extracts temporal, ocular, motion and photometric features while keeping raw frames on the phone.

03

Subjective sleepiness

A Karolinska Sleepiness Scale input adds the user's perceived state as an independent channel instead of treating it as ground truth.

04

Sleep history

Recent duration, regularity and accumulated shortfall provide longitudinal context for the current measurement.

System architecture

Quality-aware from input to action

Signal validity is checked before personalization and fusion. Missing channels are reweighted dynamically, while the resulting state feeds recovery modelling, Smart Wake and longitudinal feedback.

Read the technical architecture →
Somno system architecture diagram

Scientific foundation

Built from established sleep and fatigue science.

Somno combines measurement families used across vigilance, sleepiness, fatigue and human-performance research, then interprets them against the user's own baseline.

Vigilance

Reaction-time slowing and lapses

Psychomotor vigilance is widely used to quantify performance changes associated with sleep loss. Somno captures response speed, lapses, anticipations and variability in a compact mobile probe.

Basner, Mollicone & Dinges, 2011
Sleepiness

Karolinska Sleepiness Scale

KSS provides a fast subjective view of sleepiness and has demonstrated relationships with behavioral and EEG indicators. Somno treats it as one channel inside a broader multimodal estimate.

Kaida et al., 2006
Computer vision

Ocular fatigue markers

Eyelid closure and ocular behavior have a long research history in fatigue and vigilance monitoring. Somno derives quality-scored ocular and temporal features on-device.

Dinges et al., NHTSA
Performance

Sleep and physical readiness

Sleep loss affects cognitive and physical performance, while sleep-focused recovery strategies can support readiness. Somno makes those recovery patterns visible in a personal timeline.

Gong et al., 2024
SDI model sandbox

Inspect the fusion model.

src/engine/sdi.ts ↗

Adjust the four inputs below. The readout recomputes with the same weighting, precision scaling and missing-signal renormalization used by the app.

base 0.40
base 0.25
base 0.15
base 0.20
computed SDI 50
high confidence 4 signals
live equation 50 + 10 × 0.000 = 50
signalzweightterm
reaction0.0040.0%0.000
face0.0025.0%0.000
KSS0.0015.0%0.000
debt0.0020.0%0.000
model state

All four channels are active at full precision.

Transparent fusion

The score is inspectable.

Somno uses explicit signal weights and personal-reference normalization rather than an opaque single-output model.

SDI = clamp(50 + 10 × weighted personal-reference z-score, 0, 100)
Reaction time
40%
Face + ocular
25%
Sleep history
20%
KSS
15%

From measurement to action

Useful after the score, not just during it.

Somno connects state estimation to recovery planning, history and a native Android alarm workflow.

01

Explainable SDI

See the state estimate, confidence and strongest contributing signals instead of a black-box number.

02

Recovery modelling

Track sleep debt, recent patterns and recovery context across days rather than treating each check-in in isolation.

03

Smart Wake

A custom Kotlin alarm layer handles exact scheduling, wake flows, bounded snooze logic and reboot re-arming.

Somno result screen

Impact

Fatigue is not a student-only problem.

The same need for better fatigue awareness appears across study, sport, transport and irregular work.

ST

Students

Track accumulating sleep loss during exams and compare daily alertness with a stable personal reference.

AT

Athletes

Bring sleep shortfall, reaction-time change and perceived fatigue into a single recovery view around training, travel and competition.

DR

Drivers

Use a fast pre-task fatigue check to make changes in vigilance and recent sleep shortfall more visible before demanding journeys.

NW

Night workers

Observe alertness changes across irregular schedules, repeated short sleep and difficult circadian windows.

OC

On-call teams

Quantify the effect of disrupted nights and recovery across rotating or unpredictable duty cycles.

RS

Researchers

Explore a privacy-first mobile architecture for repeated multimodal fatigue and human-performance measurement.

Privacy by architecture

Your face stays on your phone.

Somno processes camera frames transiently on-device. The scoring pipeline works with derived numeric features rather than uploading a stream of face images to a server.

  • Raw camera frames processed locally
  • Local-first core check-ins and history
  • Optional Supabase account sync
  • Data export and account deletion controls
  • Explicit quality and availability handling
ON DEVICE Face analysis numeric features only
Camera Quality gates Features SDI

Engineering

Built like a product.

The application combines a typed React Native layer, on-device vision, local persistence, optional cloud sync and a custom native Android alarm module.

MobileReact Native + Expo + TypeScript
VisionExpo Camera + Google ML Kit
NativeKotlin Smart Wake module
StateZustand + local-first persistence
CloudSupabase auth + sync
VerificationTypecheck + logic suites + release checks

Somno v27

See your fatigue. Understand your recovery.

Explore the source, inspect the models, or install the Android build directly from GitHub.