01 — THE CHALLENGE
A useful model starts with the right question.
Most engagement metrics arrive after the session. This project asked whether subtle interaction signals could provide a useful, privacy-conscious indication much earlier.
03 — Behavioral ML · Real-time inference
CursorAI
Predict engagement in the first seconds.
Try the live demo↗PROJECT OVERVIEW
A behavioral machine-learning pipeline that reads early mouse dynamics to estimate user engagement while there is still time to adapt the experience.
Project anatomy
From micro-movements to an early engagement score.
The pipeline aggregates session events, extracts spatial and temporal features, then applies a low-latency XGBoost model.
Core pipeline
Select a step to understand how the data is transformed.Selected step
The frontend captures cursor positions, timestamps, clicks and scrolling while the interface is explored.
Implemented features
Visible capabilities are connected to the technical components that make them possible.Live trajectory
The interface visualizes movements actually produced during the demonstration.
Multi-batch session
Multiple interaction fragments are combined before inference.
Explainable features
Velocity, acceleration, distance and direction changes remain visible and interpretable.
Early prediction
The score is computed during the session, before a final conversion or exit.
Decision threshold
The probability is translated into low or high engagement through an explicit rule.
Storage with local fallback
Sessions use Google Cloud Storage when available and the local filesystem otherwise.
Stack by responsibility
Each technology is placed within its role in the system.React
Experience instrumentation and score presentation.
FastAPI
Batch collection and session prediction endpoint.
Pandas · NumPy
Trajectory reconstruction and normalization.
Feature engineering
Kinematics, direction, distance and temporal aggregates.
XGBoost · Joblib
Engagement probability and model serialization.
GCS · Cloud Run
Session persistence and API execution.
Real-time behavioral inferenceBased on the current codebase · Simplified representation
From technology to use
Generate the signals analyzed by the model yourself.
Move your mouse, pause briefly and explore the interface. Trajectory, speed and direction changes progressively feed the behavioral interpretation.
Educational demonstration: data is local and some computations are simulated or accelerated.04 — OUTCOME AND IMPACT
“An actionable early-session score that can support personalization before conventional conversion signals become available.”