Google AI Classroom Part 2: Zero-Egress Browser Edge AI Investigator — Running Gemma 4 E2B, MediaPipe 3D Face Mesh & LiteRT Whisper in Web Workers

Author: Cyrus Wong (黃俊彥) — Google Developer Expert - Cloud (AI), Senior Lecturer at HKIIT

1. Problem Statement: Why Proctored Tests & Practical Examinations Need an On-Device Browser Edge AI Investigator

In proctored tests and practical examinations, student screen captures, webcam frames, and audio evidence are recorded and uploaded to Cloud Storage for official exam invigilation and forensic audit trails. However, relying solely on human invigilators or cloud-polled snapshots during a live 40-to-100+ student examination creates three critical operational blind spots:

1. The Human Invigilator Attention Bottleneck: A single lecturer or invigilator cannot simultaneously watch 40+ student screens, 40+ live webcam feeds, and listen to 40+ microphones in real time to catch fleeting exam misconduct—such as glancing at a concealed smartphone, peeking at a neighbor's monitor, whispering answers, or querying a voice assistant.
2. Fleeting Violations Between Snapshots vs. 432,000+ Cloud AI Calls: Periodic cloud screen and webcam snapshots provide the historical archive, but a 3.5-second glance at a cheat sheet or a whispered answer often occurs between polling intervals. Conversely, streaming 30-FPS video and continuous audio from 40 exam seats to Cloud LLMs for real-time AI inspection every second would require 432,000+ cloud API calls per 3-hour exam—saturating exam lab bandwidth, triggering API rate limits, and incurring runaway cloud inference costs.
3. High False-Positive Rates Without Local Geometric & Semantic Reasoning: During a practical coding or cloud exam, students naturally blink, glance down at their keyboard, read exam questions aloud to themselves ("Let me check Question 3"), or call the invigilator for technical assistance ("Mr. Wong, my Docker daemon crashed"). Naive motion or audio-level triggers flood invigilators with false alarms.

Using Google MediaPipe (@mediapipe/tasks-vision), Google LiteRT.js (@litertjs/core Whisper-Tiny), and LiteRT-LM WebGPU (@litert-lm/core Gemma 4 E2B gemma-4-E2B-it-web.litertlm), we turned every student's browser into an always-on Zero-Egress Browser Edge AI Investigator—continuously inspecting 3D head pose, iris metric distance, and spoken semantic intent at 60 FPS with $0.00 cloud AI inference egress, and immediately flagging confirmed exam irregularities with synchronized dual screen + webcam + audio evidence for the invigilator.

2. Architectural Deep-Dive: Off-Main-Thread Web Worker & Hardware Frame Sync Pipeline

Running a 478-point 3D facial landmark detector, an iris-depth estimator, a Chirp-Z log-Mel spectrogram extractor, a Whisper speech-to-text model, and a 2-billion-parameter Gemma 4 E2B LLM on the main React thread would freeze the browser UI immediately.

In useFaceMonitor.js, litertWhisper.worker.js, litertGemma.worker.js, and useClientLiteRTGemma.js, we isolate speech and LLM reasoning inside dedicated Web Workers while synchronizing vision landmark detection to the browser's hardware video decoder via requestVideoFrameCallback():

3. Implementation Walkthrough 1: 478-Point 3D Face Mesh, Iris Metric Depth & The 3-Second Debounce State Machine

Legacy proctoring software is notorious for false-positive alerts: if a student sneezes, blinks, or glances down at their mechanical keyboard for 500 milliseconds, they get flagged for "cheating." Conversely, if a student sits in a corner desk where their webcam is mounted at a $15^\circ$ angle, uncalibrated head-pose detectors flag them continuously.

We solved both problems in useFaceMonitor.js through three mathematical and state-machine innovations:

1. 1-Click Adaptive Neutral Pose Calibration (calibrateBaseline): Students click 🎯 Set Center Pose to zero out their personal baseline head Yaw ($\text{Yaw}_0$) and Pitch ($\text{Pitch}_0$) offsets for off-center monitor mounts.
2. Monocular Depth-from-Iris Distance Estimation ($D_{\text{cm}}$): Using the anatomical constancy of the human horizontal iris diameter ($d_{\text{iris}} \approx 11.7\text{ mm}$) and MediaPipe's iris boundary landmarks (469, 471 for left iris; 474, 476 for right iris), we compute physical camera-to-face distance in centimeters without any hardware depth sensor:

1. 3-Second Debounce Gate with 600ms Blink/Jitter Grace Window & Lifecycle Incident Tracking: An anomaly (looking_away, no_face, multiple_faces, eyes_closed, talking) must persist continuously for debounceSeconds (default 3.0 seconds) before opening a single active incident document in Firestore (irregularities/{id}), and automatically updates endedAt and durationSeconds when the student returns to normal posture!

Production Code Listing 1: 3D Iris Depth, Eye-Gaze Ratio & Anti-Jitter Debounce in web-app/src/hooks/useFaceMonitor.js

Here is the exact real-time biometric geometry and debounce gate running inside useFaceMonitor.js:

// MediaPipe Iris: Depth-from-Iris Distance Calculation (11.7mm constant human iris diameter) let calculatedDistanceCm = 55; if (landmarks[469] && landmarks[471] && landmarks[474] && landmarks[476]) { const leftIrisPx = Math.hypot(landmarks[469].x - landmarks[471].x, landmarks[469].y - landmarks[471].y) * (video.videoWidth || 640); const rightIrisPx = Math.hypot(landmarks[474].x - landmarks[476].x, landmarks[474].y - landmarks[476].y) * (video.videoWidth || 640); const avgIrisPx = (leftIrisPx + rightIrisPx) / 2; if (avgIrisPx > 2) { const focalLength = 0.85 * (video.videoWidth || 640); calculatedDistanceCm = Math.round((focalLength * 1.17) / avgIrisPx); calculatedDistanceCm = Math.max(15, Math.min(220, calculatedDistanceCm)); setMetricDistance(calculatedDistanceCm); } } // Horizontal Yaw & Vertical Pitch with Adaptive Neutral Baseline Calibration Offset const dLeft = Math.hypot(nose.x - leftEyeOuter.x, nose.y - leftEyeOuter.y); const dRight = Math.hypot(nose.x - rightEyeOuter.x, nose.y - rightEyeOuter.y); const rawYaw = Math.round(((dLeft - dRight) / (dLeft + dRight)) * 100); const dNoseChin = chin.y - nose.y; const dNoseForehead = nose.y - forehead.y; const rawPitch = Math.round(((dNoseChin - dNoseForehead) / (dNoseChin + dNoseForehead)) * 100); const calculatedYaw = rawYaw - (baselineOffsetRef.current?.yaw || 0); const calculatedPitch = rawPitch - (baselineOffsetRef.current?.pitch || 0); // Anti-Jitter Debounce Gate Logic with 600ms Grace Window const now = Date.now(); if (currentAnomaly) { lastAnomalyTimeRef.current = now; if (!anomalyStartRef.current || anomalyStartRef.current.type !== currentAnomaly) { anomalyStartRef.current = { type: currentAnomaly, startTime: now, details: detailsMessage }; } else { const elapsedSeconds = (now - anomalyStartRef.current.startTime) / 1000; if (elapsedSeconds >= debounceSeconds) { setActiveViolation(currentAnomaly); startIncident(currentAnomaly, detailsMessage); } } } else { // Grace window of 600ms to tolerate natural blinks or momentary landmark noise const timeSinceLastAnomaly = now - (lastAnomalyTimeRef.current || 0); if (timeSinceLastAnomaly > 600) { anomalyStartRef.current = null; setActiveViolation(null); if (activeIncidentDocRef.current) { resolveIncident(); } } } 

4. Implementation Walkthrough 2: On-Device Whisper-Tiny Log-Mel DSP & Special Language Token Decoding

In Hong Kong classrooms, students and lecturers naturally code-switch between Cantonese, Mandarin, and English technical terminology within a single sentence (e.g., "唔該阿Sir，呢個 Question 3 個 option B 係咪啱㗎？").

To transcribe student speech on-device with zero UI jank, litertWhisper.worker.js and whisperTokenizer.js run @litertjs/core inside a dedicated Web Worker:

1. Dynamic Peak Auto-Gain & 80-Bin Log-Mel Spectrogram via Chirp-Z FFT: Whisper expects a normalized [1, 80, 3000] Float32 tensor representing 30 seconds of 16 kHz audio. Because quiet whispering during an exam often has low amplitude ($0.003 < \text{peak} < 0.25$), extractAudioFeatures() applies up to $6\times$ automatic gain normalization before computing a 400-point Hann-windowed Chirp-Z Fast Fourier Transform (hopLength = 160) and projecting onto 80 Slaney-normalized Mel filterbanks.
2. Whisper Special Language Token Decoding & WebGPU-to-WASM Delegate Failover: Rather than relying on brittle regex heuristics, extractWhisperLanguageToken() reads Whisper's native special language token (50259..50357, e.g. <|en|>, <|zh|>, <|yue|>) directly from the decoded token stream right after <|startoftranscript|> (50258), while Gemma 4 E2B (litertGemma.worker.js) and Cloud Gemini (analyzeAudioFlow) perform semantic language and code-switching classification ("language": "cantonese|mandarin|english|mixed") alongside exam intent evaluation.

Production Code Listing 2: WebGPU/WASM Whisper Compilation & Special Language Token Decoder in litertWhisper.worker.js and whisperTokenizer.js

/** * Extracts the language code emitted by Whisper's special language token * (`<|en|>`, `<|zh|>`, `<|yue|>`, etc.) in the output token sequence. */ export function extractWhisperLanguageToken(tokenIds, vocabulary = [], fallbackLanguage = 'auto') { if (!tokenIds || typeof tokenIds === 'string') return fallbackLanguage; const prefixLength = Math.min(tokenIds.length || 0, 8); for (let i = 0; i < prefixLength; i++) { const tokenId = Math.round(tokenIds[i]); const vocabToken = vocabulary?.[tokenId]; if (typeof vocabToken === 'string') { const match = vocabToken.match(/^<\|([a-z]{2,3})\|>$/i); if (match) { return match[1].toLowerCase(); } } if (tokenId >= WHISPER_LANGUAGE_TOKEN_START && tokenId <= WHISPER_LANGUAGE_TOKEN_END) { return WHISPER_LANGUAGE_ID_MAP[tokenId] || fallbackLanguage; } } return fallbackLanguage; } export async function compileWhisperModel(loadAndCompile, modelBuffer, preferredDelegate) { const compile = (accelerator) => loadAndCompile( new Uint8Array(modelBuffer), { accelerator } ); if (preferredDelegate !== 'webgpu') { return { model: await compile('wasm'), delegate: 'wasm' }; } try { return { model: await compile('webgpu'), delegate: 'webgpu' }; } catch (webGpuError) { console.warn( '[LiteRTWorker] WebGPU could not compile the dynamic Whisper graph; retrying with WASM:', webGpuError?.message || webGpuError ); return { model: await compile('wasm'), delegate: 'wasm' }; } } 

5. Implementation Walkthrough 3: On-Device Gemma 4 E2B (gemma-4-E2B-it-web.litertlm) for Semantic Proctoring, Dialect Classification & Live Translation

Recognizing speech tokens is only half the battle: if a student mutters "Let me read this question carefully" (benign self-talk) or raises their hand and says "Mr. Wong, my Docker daemon crashed" (legitimate inquiry), a keyword-based detector would falsely accuse them of cheating.

We needed a true Large Language Model running locally in the browser to evaluate semantic intent and spoken dialect/code-switching. Using @litert-lm/core (v0.15.0) and the quantized Gemma 4 E2B model (gemma-4-E2B-it-web.litertlm, ~1.91 GB in WebGPU), litertGemma.worker.js serves two mission-critical classroom roles at $0.00 API cost:

1. EVALUATE_TRANSCRIPT (Exam Acoustic Intent & Dialect Investigator): Classifies student speech transcripts into one of 5 strict categories (COLLUSION_EXAM, EXTERNAL_AI_ASSIST, UNAUTHORIZED_TALK, LEGITIMATE_INQUIRY, or BENIGN) along with spoken dialect/code-switching ("language": "cantonese|mandarin|english|mixed"), severity, confidence, quoted evidence, and rationale—and supports custom teacher prompts ({{transcript}}, {{classId}}, {{studentEmail}}) authored in the AI Prompt Studio!
2. TRANSLATE_TRANSCRIPT ($0.00 Client-Mode Live Subtitles): When a lecturer selects On-Device Gemma 4 E2B in the Live Subtitles modal, the same WebGPU worker translates Cantonese/English code-switched lecture speech into up to 7 target languages (en, zh-Hant, zh-Hans, ja, ko, es, fr) while preserving technical identifiers like useEffect, Docker, and Python.

Production Code Listing 3: Gemma 4 E2B WebGPU Initialization & Structured Output Parser in web-app/src/workers/litertGemma.worker.js

const GEMMA_OUTPUT_INSTRUCTIONS = `Use exactly one category: - COLLUSION_EXAM: asking for, offering, or discussing exam answers, questions, or options - EXTERNAL_AI_ASSIST: asking a voice assistant, search engine, phone, or AI tool for help - UNAUTHORIZED_TALK: unrelated conversation with another person during a silent exam - LEGITIMATE_INQUIRY: procedural or technical help requested from the teacher or proctor - BENIGN: self-talk, reading aloud, ambient speech, coughing, or silence Also classify the spoken language/code-switching of the transcript in "language" as one of: "cantonese", "mandarin", "english", or "mixed". Respond with only one JSON object and no markdown: {"isViolation":boolean,"category":"COLLUSION_EXAM|EXTERNAL_AI_ASSIST|UNAUTHORIZED_TALK|LEGITIMATE_INQUIRY|BENIGN","severity":"critical|high|medium|low|none","confidence":number,"language":"cantonese|mandarin|english|mixed","evidence":"quoted key phrase","rationale":"short explanation"}`; async function initializeGemma(modelUrl) { if (gemmaEngine) return { fromCache: gemmaLoadedFromCache }; if (initializationPromise) return initializationPromise; initializationPromise = (async () => { const ua = self.navigator?.userAgent || ''; const isMobileOrIos = /iPhone|iPad|iPod|Android|CriOS|FxiOS/i.test(ua) || (self.navigator?.platform === 'MacIntel' && Number(self.navigator?.maxTouchPoints) > 1); if (isMobileOrIos) { throw new Error( 'On-device Gemma 4 E2B (~1.9 GB) exceeds mobile browser memory limits; using Cloud fallback.' ); } if (typeof self.navigator?.deviceMemory === 'number' && self.navigator.deviceMemory < 4) { throw new Error( `Device RAM (${self.navigator.deviceMemory} GB) is below the 4 GB minimum for Gemma 4 E2B; using Cloud fallback.` ); } if (!navigator.gpu) { throw new Error('WebGPU is unavailable on this device'); } configureLiteRtLmWasmLocation(); const { modelStream, fromCache, cacheWritePromise } = await streamModelWithProgress(modelUrl); self.postMessage({ type: 'STATUS', payload: { status: 'loading', progress: 90 } }); gemmaEngine = await Engine.create({ model: modelStream, mainExecutorSettings: { maxNumTokens: 2048 }, }, GEMMA_PROCTOR_SYSTEM_PROMPT); const cached = await cacheWritePromise; gemmaLoadedFromCache = fromCache; return { fromCache, cached: fromCache || cached }; })(); try { return await initializationPromise; } finally { initializationPromise = null; } } 

6. Edge Cases & Production Failure Modes We Solved: Fresh "DeepFreeze" Lab PCs, Dynamic Tensor Guards, iOS Jetsam OOM & Live Hot-Swapping

Deploying a 1.91 GB WebGPU LLM (gemma-4-E2B-it-web.litertlm) and an on-device Whisper TFLite worker across university computer labs and BYOD student devices exposed four real-world failure modes that we engineered around in StudentView.jsx, gemmaLiteRTLoader.js, useClientLiteRTGemma.js, and litertWhisper.worker.js:

1. "DeepFreeze" Wiped Lab PCs & Automatic Uncached Pre-Warming: University lab PCs reboot into a clean image every morning, wiping CacheStorage (isGemmaCached === false). If auto-preloading gated on isGemmaCached, a wiped lab PC would stay stuck at gemmaStatus: 'idle' (gemmaEngine: 'uninitialized') forever unless the student manually clicked a preload button. Instead, StudentView.jsx automatically triggers preloadGemmaModel() whenever preloadClientAi is broadcast by the teacher, whenever active audio monitoring starts (isAudioCaptureActive), or on the very first spoken transcript (handleLiveVoiceTranscript), requesting navigator.storage.persist() and streaming the 1.91 GB .litertlm bundle directly into Engine.create({ model: modelStream }).
2. Non-Blocking Cloud-to-Edge Live Hot-Swap: If an exam begins while Gemma 4 E2B is still downloading on a wiped lab PC, the student is never blocked from starting their exam. Any speech detected while isGemmaReady is false is transparently routed to our Cloud Genkit flow (analyzeAudio using gemini-3.5-transcribe-preview + gemini-3.5-flash-lite), and automatically hot-swaps to 100% local Gemma 4 E2B WebGPU inference the millisecond the worker emits INIT_COMPLETE (ready: true)!
3. Dynamic-Output TFLite Tensor Allocation Guard (clientInferenceSupported): Certain exported Whisper TFLite graphs declare their output sequences tensor with kTfLiteDynamic allocation, which causes @litertjs/core WASM LiteRtCompiledModel.run() to throw Failed to register output tensor buffer: Failed to set custom allocation for output. Our worker catches this on the first segment and flips clientInferenceSupported = false to prevent repeated WASM C++ error spam while seamlessly continuing transcription via the bound MediaStreamTrack recognizer and Firebase AI Logic STT fallback.
4. iOS / iPadOS WebKit Jetsam OOM Guard & Storage Quota Pre-Flight: Allocating a 1.91 GB WebGPU buffer on an iPhone or iPad (which masquerades as MacIntel with maxTouchPoints > 1 on iPadOS) triggers Apple's kernel Jetsam killer, crashing the browser tab. Similarly, laptops with < 2.5 GB free disk space throw QuotaExceededError mid-download. Our isMobileOrIosBrowser(), deviceMemory < 4, and precheckGemmaViability() guards fail fast in < 5ms and route mobile/low-RAM devices cleanly to Cloud Fallback.

Production Code Listing 4: Cloud-to-Edge Hot-Swap & Dual Firestore Alerting in StudentView.jsx and useClientLiteRTGemma.js

Notice how StudentView.jsx automatically starts preloadGemmaModel() if gemmaStatus === 'idle' while routing transcripts to Cloud Genkit until isGemmaReady becomes true, and how useClientLiteRTGemma.js writes confirmed violations and LLM-classified speechLanguage to Firestore:

Related posts

La fin de la Quest 2 et ce qu'elle dit vraiment de la VR comme outil d'entraînement visuel

La fin de la Quest 2 et ce qu'elle dit vraiment de la VR comme outil d'entraînement visuel Meta vient d'annoncer la fin

0dev.to News

Tennessee Layoffs November 2026: WARN Numbers Explained

Tennessee Layoffs November 2026: WARN Numbers Explained Tennessee's 2026 layoff tally has not moved in a month: the Layo

0dev.to News

Building a Product Page for French Cotton Bedding: 6 Things That Moved Conversions

I run a small bedding store, Cozy Home Textiles ( https://www.webcozyhometextiles.com/ ). We sell French-style cotton du

0dev.to Design

What Multi-User Adds to a VR Project: Scope, Skills and Risk

What changes when a VR project becomes multi-user: authority, latency, identity, voice, permissions and backend, with a

0dev.to Discuss

Context-Aware Dictation: Why Voice Typing Should Sound Different in Slack, Outlook and VS Code

Voice typing has a quiet problem. The words you say are the same, but the place you say them changes what good text look

0dev.to Productivity

108 free websites and tools for IT and AI practitioners, checked one by one.

I put this list together for people who are tired of paying for subscriptions when a free tool already does the job well

0dev.to Productivity