CLIO — Cognitive Load Inference System

CLIO is a local, privacy-preserving system for inferring cognitive load during focused work sessions using behavioral interaction data.
The system models how a user interacts with their computer over time and estimates mental states such as focus, strain, and fragmentation by comparing live behavior against a continuously evolving personal baseline.

This project is designed as a real-world applied intelligence system, not a synthetic demonstration or offline ML experiment.

---

Problem Statement

Most productivity and focus-tracking tools rely on self-reports, fixed heuristics, or cloud-based monitoring. These approaches are often subjective, intrusive, or opaque.

CLIO addresses this by:

• Inferring cognitive load from natural interaction behavior
• Running entirely on-device
• Avoiding content capture
• Learning personalized baselines instead of population averages

---

System Overview

CLIO consists of two decoupled components:

1. Backend Cognitive Load Engine

• Captures behavioral signals during a session
• Computes cognitive load snapshots at fixed intervals
• Updates a rolling personal baseline
• Stores structured session data locally

2. Frontend Analytics Dashboard

• Visualizes the current session only
• Displays trends, confidence, regimes, and baseline context
• Reads data produced by the backend (read-only)

The frontend never controls sensors or processes raw input.

---

Behavioral Signals Used

CLIO infers cognitive load from the following signals:

• Keyboard dynamics
  • Typing rate
  • Typing variance
  • Error behavior
• Mouse behavior
  • Activity volume
  • Movement entropy
• Optional experimental signals
  • Camera engagement (binary presence only)
  • Microphone engagement (binary presence only)

No keystroke content, screen data, audio, or video is recorded or stored.

---

Cognitive Load Modeling

CLIO uses deviation-based inference rather than supervised classification.

For each snapshot:

1. Behavioral features are normalized against a rolling personal baseline
2. Feature deviations are aggregated into a Cognitive Load Index (CLI)
3. A confidence score reflects signal consistency
4. A regime classifier assigns a cognitive state:
   • focused
   • strained
   • fragmented

Baselines evolve across sessions, allowing the model to adapt to the individual user over time.

---

Outputs

For each session, CLIO generates:

• Cognitive Load Index (CLI) over time
• Model confidence over time
• Cognitive regime distribution
• Session-level summary statistics
• Historical baseline context aggregated across past sessions

---

Privacy and Design Principles

• All computation runs locally
• No cloud services
• No raw sensor streams stored
• No personal content captured
• Camera and microphone usage is optional and minimal
• Session data can be deleted or disabled entirely

The system is fully inspectable and auditable.

---

Technology Stack

• Python
• Streamlit
• NumPy
• Pandas
• Plotly
• Pynput

---

Running the System

Backend: Run a Cognitive Load Session

python main.py --duration 600 --interval 20

Optional Flags

--camera   # Enable camera engagement signal (experimental)
--mic      # Enable microphone engagement signal (experimental)

Example: Backend Session with Optional Signals Enabled

python main.py --duration 600 --interval 20 --camera --mic 

Frontend: Launch the Dashboard (Read-Only)

streamlit run ui/dashboard/app.py