How it works

Simple to wear.
Intelligent inside.

You put on the band, open the app, and study. StudyTwin reads your body, calculates your state, and adapts your session — automatically.

Signal processing pipeline

Five layers.
Distributed intelligence.

ML inference at the edge. Sync in the cloud. Adaptation in the browser. Pre-analysis in the GPU cluster.

L1
Sensing
Wearable · Physical
Biosignal Acquisition (MAX30102/GSR)
MAX30102 (PPG) · GSR Electrodes · MediaPipe

Three validated biomarkers of cognitive load, captured in real time. The Grove GSR sensor outputs a signal conditioned by an LM324 op-amp. MAX30102 extracts inter-beat intervals via 880nm infrared PPG for HRV. Blink rate runs in-browser via MediaPipe Face Mesh — zero extra hardware.

MAX30102 · 880nm PPG Grove GSR · LM324 I2C · ADC GPIO34 MediaPipe Face Mesh 60s baseline calibration
L2
Edge AI
ESP32 · On-device
1D-CNN Inference (Edge AI)
1D-CNN TinyML Model · 47ms Inference Latency

The ESP32 WROOM-32 runs a 1D-CNN trained on Edge Impulse, performing 4-class cognitive state inference entirely on-device at 47ms latency. Signal preprocessing — EMA smoothing (alpha = 0.28), baseline drift correction, RMSSD extraction — runs in Arduino C++ firmware.

ESP32 WROOM-32 · 240MHz 1D-CNN · Edge Impulse EMA smoothing · alpha=0.28 47ms inference SSD1306 OLED display
L3
Cloud Sync
Firebase · Cloud
Synchronisation Layer
Realtime DB · Auth · Firestore

Firebase Realtime Database receives the ESP32 JSON payload and immediately pushes it to all subscribed browser clients via onSnapshot(). Mean propagation latency: 154ms. Firebase Auth handles per-user isolation. Firestore stores session history for longitudinal analysis.

Firebase Realtime DB onSnapshot() push 154ms propagation Firebase Auth · multi-user
L4
Dashboard
Browser · PWA
Closed-Loop Adaptation
Three.js · Chart.js · Adaptive Pomodoro

The dashboard fuses the three biosignals into a final CLI score, drives the adaptive Pomodoro engine, renders the 3D Digital Twin, and generates session insights. All adaptation logic runs client-side — it functions even if Firebase degrades.

3D Neural Twin · WebGL Chart.js · rolling CLI Adaptive Pomodoro engine AI insights generation
L5
TRIBE v2
Kaggle GPU · Pre-session
Brain Circuit Intelligence
Meta TRIBE v2 · fMRI Foundation Model

Before each session, TRIBE v2 — Meta's brain-predictive foundation model trained on 1,115 hours of fMRI data from 720 subjects — analyses your study material and predicts which brain circuits it will demand. Content Demand Scores for three regions (executive, language, visual) are written to Firebase session metadata. The adaptive engine uses these scores to set personalised session parameters before you put on the headband.

20,484 cortical vertices 3 Content Demand Scores Kaggle T4 GPU · Free tier Pre-session analysis
The signals

Three ways your body
tells us how you're doing.

Skin conductance (GSR)

When your brain works hard, your sympathetic nervous system activates and your skin conducts electricity differently. The Grove GSR sensor on your wrist detects this in real time — a reliable, well-studied marker of cognitive load.

50% weight

Heart rate variability (HRV)

The variation between heartbeats is deeply meaningful. High variability means your body is calm. When HRV drops, your nervous system is under load. MAX30102 measures this continuously via 880nm infrared PPG.

35% weight

Blink rate

Normal: 15-20 blinks per minute. Deep focus or fatigue drops this below 10. Your webcam detects this via MediaPipe Face Mesh — no hardware, no clip, nothing to attach to your face.

15% weight
Weighted fusion (IEEE §V)

One equation. One score.

Cognitive Load Index

CLI = (GSR × 0.50) + (HRV × 0.35) + (Blink × 0.15)

Each signal is normalised to 0-100, then weighted according to predictive power validated against NASA-TLX ground truth. Smoothed with EMA (alpha = 0.28) to prevent jitter.

Your states

One score. Four states.
A different response to each.

Your Cognitive Load runs 0-100. StudyTwin maps it to four states and responds differently to each.

0-25
Calm

Your body is relaxed and your mind has headroom. A great window for complex, new material.

→ Timer extended +5 min
→ Harder material suggested
26-55
Focused

Engaged and handling load well. You're in the optimal learning zone — no changes needed.

→ Timer runs normally
→ No intervention
56-77
Elevated

Load is climbing. Still functional, but retention is dipping. Time to ease up slightly.

→ Session shortened -4 min
→ Gentler task suggested
78-100
Overloaded

Working memory saturated. Continuing now costs more than it gives. A break restores capacity.

→ Break triggered (15 min)
→ Timer paused
TRIBE v2 Intelligence

How TRIBE v2 Changes
the Routing Logic.

1
Material uploaded to Kaggle GPU
2
TRIBE v2 predicts circuit demand
3
CDS scores written to Firebase
4
Dashboard reads session metadata
5
Session parameters personalised
6
Biosensors take over in real time
Executive CircuitHigh Demand

Loaded by complex logic, multi-step problem solving, and abstract reasoning. Demands high working memory.

→ ROUTE TO: Visual review material (diagrams, flowcharts)
Language CircuitHigh Demand

Loaded by dense reading, vocabulary acquisition, and verbal reasoning. Saturated by technical prose.

→ ROUTE TO: Numerical or Visual content (equations, graphs)
Visual CircuitHigh Demand

Loaded by complex animations, spatial reasoning, and diagram-heavy content. Saturated by high-motion video.

→ ROUTE TO: Audio or Text-light reading material
Data pipeline

Signal to action, step by step.

Step 01 · Sense
Capture
  • → GSR ADC reading
  • → MAX30102 IBI
  • → MediaPipe blink
  • → SpO2 percentage
Step 02 · Process
Edge AI
  • → 60s calibration
  • → EMA smoothing
  • → RMSSD extraction
  • → 1D-CNN at 47ms
Step 03 · Fuse
Fusion
  • → CLI weighted sum
  • → JSON payload built
  • → Firebase push
  • → 154ms propagation
Step 04 · Adapt
Closed-Loop
  • → Pomodoro mutation
  • → Timer shortening
  • → Neural twin shift
  • → Session logging