HawkWatch
AI-Powered Proctoring & Computer Vision Surveillance
01 / Project Overview
A browser-based AI-powered proctoring and real-time student monitoring platform. Integrating light-weight computer vision models directly in the client browser, HawkWatch tracks eye movements, head orientation, multi-face presence, and audio triggers, securing online examinations without heavy desktop installations.
02 / The Challenge & Problem
Real-World Problem Statement
Online testing is plagued by academic integrity concerns. Traditional proctoring tools require invasive desktop applications that collect massive system data, raise privacy concerns, and suffer from poor cross-platform compatibility on older student laptops.
03 / The Engineering Solution
Implementation & Architectural Approach
Designed a browser-only proctoring solution using MediaPipe and TensorFlow.js. All computer vision inference (such as facial landmark tracking and gaze estimation) is executed on the student's local machine, ensuring 100% data privacy and reducing server compute costs.
04 / Technical Architecture Flow
Browser WebAssembly & MediaPipe
Processes student camera feed on the local GPU/WASM threads, outputting facial mesh coordinates.
Offscreen Canvas Watchdog
Evaluates face angles, eye tracking lines, and ambient voice audio peaks in background workers.
React Control Room UI
Feeds incident logs, live statistics, and alert highlights to examiners via WebSockets.
05 / Key Project Features
Local Gaze Tracking
Uses local face landmark coordinates to detect if the student shifts sight away from the screen.
Multi-Face Watchdog
Flags warning signals immediately if secondary individuals enter the camera's viewport.
Audio Anomaly Logger
Analyzes microphone feeds to detect human speaking patterns and sudden volume spikes.
06 / Engineering Challenges & Mitigations
Running computer vision models concurrently with client exam questions caused lag on entry-level computers.
Delegated heavy model execution to Web Workers using OffscreenCanvas, ensuring the UI main thread stays locked at 60fps.
Students shifting positions slightly were incorrectly flagged with cheating violations.
Created calibration routines before exams, adjusting baseline head coordinates and gaze boundaries.
07 / Technical & Personal Learnings
Acquired key insights in browser-based model quantization, WebAssembly bindings, and off-thread model inference.
Gained deep experience in asynchronous audio-visual stream processing and real-time Socket communication.