SHADOWNOTES // FIELD MANUAL

Project ShadowNotes v1.3.1

Classification On-Device AI / WebAssembly

Network Status AIR-GAPPED (Post-Download)

Hackathon RunAnywhere Vibe Challenge

Vision Atmanirbhar Bharat / #India2047

Team Divya Mohan & Kumkum Thakur

Institution Shoolini University

Facilitated By ThoughtWorks Technologies

▸ FIELD OPERATIONS // QUICK START

Access the Application

Navigate to shadownotes.dmj.one. The boot sequence initializes the secure environment, registers the LLM model backend, and downloads the on-device AI engine (~810 MB, one-time).

First load takes 1-3 minutes depending on connection. After that, the app works fully offline.

Select Operational Domain

Choose one of 4 domain profiles: Security Audit, Legal Deposition, Medical Notes, or Incident Report. Each domain has its own AI extraction pipeline and intelligence categories.

Begin Capture Session

Click BEGIN CAPTURE SESSION to enter the active capture screen. You'll see the case number, timer, and the split-view layout with transcript on the left and intelligence on the right.

Choose Capture Mode

Three capture modes are available: MIC — Push-to-talk voice capture with real-time transcription. TEXT — Manual text input for quiet environments. AGENT — Hands-free VoiceAgent mode using VoicePipeline + VoiceAgent for continuous listen-process-respond without manual button presses.

Try: "Patient presents with elevated blood pressure, 160 over 95. Prescribing Telmisartan 40mg daily." Then click DECODE INTELLIGENCE to extract findings.

Decode & Review Intelligence

Click DECODE INTELLIGENCE to trigger AI extraction. The streaming LLM (TextGeneration.generateStream()) processes your transcript token-by-token, showing results in real-time. StructuredOutput.extractJson() validates the extraction format. All domain categories are visible — click + ADD on any category to manually add findings.

Click any extracted finding to edit it. Click the X to remove incorrect entries. On production hardware with GPU, extraction can run live.

End Session & Review Dossier

Click END SESSION to view the complete session dossier with all transcripts and intelligence findings. You can still edit findings on this screen.

Destroy All Data

Click DESTROY SESSION and confirm. A cinematic burn animation plays as all data is wiped from browser memory and encrypted vault. Zero trace remains.

▸ OPERATIONAL DOMAINS

🛡 Security Audit — Kavach

OPERATION FIREWALL • “The Divine Shield”

TOP SECRET

Extracts: Vulnerabilities, Timeline, Evidence, Affected Systems, Risk Assessment

⚖ Legal Deposition — Nyaaya

OPERATION TESTIMONY • “The Path of Justice”

CONFIDENTIAL

Extracts: Key Statements, Timeline, Parties Involved, Contradictions, Exhibits

🩺 Medical Notes — Sanjeevani

OPERATION VITALS • “The Life-Giving Herb”

RESTRICTED

Extracts: Symptoms, Diagnoses, Medications, Vital Signs, Follow-up Actions

🚨 Incident Report — Prahari

OPERATION CHRONICLE • “The Vigilant Sentinel”

SECRET

Extracts: Incident Timeline, Witnesses, Damage Assessment, Root Cause, Next Steps

▸ VERIFICATION PROTOCOL // TESTING

☐ Browser requirement: Use Chrome, Edge, or Safari. Firefox does not support Web Speech API for live transcription.

☐ Microphone access: Allow microphone when prompted. Without it, speech capture will not function.

☐ AI model download: Wait for the boot screen to show the green checkmarks. The ~810 MB model (Gemma 3 1B Instruct) downloads once and is cached in the browser's private filesystem (OPFS).

☐ Offline test: After first load, disconnect from the internet entirely. Open the app again — it should boot and run the AI model completely offline. Open DevTools Network tab to verify zero outgoing requests during capture.

☐ Privacy test: During a capture session, check DevTools Network tab. Zero requests should be made. All AI processing happens on your device.

☐ Data destruction: After destroying a session, check DevTools Application tab — the encrypted vault entry for that case is removed. OPFS model cache (AI weight file) remains for reuse.

☐ Edit intelligence: Click on any extracted finding to edit it inline. Press Enter to save, Escape to cancel. Click the X button to remove incorrect findings.

▸ CAPABILITIES

🎙

Browser-Native Transcription

Web Speech API provides fast, lightweight speech-to-text that works on any device with zero model download.

🧠

Streaming AI Extraction

Gemma 3 1B Instruct LLM streams tokens in real-time via TextGeneration.generateStream(). StructuredOutput.extractJson() validates extraction format. Advanced sampling with topK, topP, temperature, and stopSequences.

📝

Keyword Fallback

Instant regex-based extraction while the LLM loads, ensuring intelligence findings from the first word spoken.

🔒

Encrypted Vault

AES-256-GCM encryption with per-case HKDF-derived keys. WebAuthn biometric auth (Windows Hello / Touch ID). Encrypted export/import via .shadow files.

✏

Editable Findings

Click any AI-extracted finding to correct it inline. Remove incorrect entries with one click. Voice commands: "Hey Shadow, delete case..."

🔥

Session Destruction

Animated data destruction with phase-by-phase memory wipe. Encrypted vault entry removed. Zero trace remains.

🔊

VoiceAgent Mode

Hands-free operation via VoicePipeline + VoiceAgent. Continuous listen → process → respond loop. No button presses needed — speak and the AI extracts intelligence automatically.

🧩

RAG Context Injection

Prior findings are semantically retrieved via embeddings and injected into the LLM prompt. Extraction accuracy improves across sessions as context builds.

🛡

Brute-Force Protection

Vault unlock is rate-limited with exponential backoff (5s / 15s / 30s / 60s) after 3 failed attempts. Prevents passphrase guessing attacks.

🔍

Semantic Search

Global search results are reranked by semantic relevance using embeddings-based findSimilar(). Find intelligence by meaning, not just keywords.

⚙

Deep SDK Integration

RunAnywhere SDK: All 3 packages (web, web-llamacpp, web-onnx) — 20+ genuinely load-bearing features including LLM streaming, ToolCalling, Embeddings, STT, TTS, VAD, AudioCapture, and AudioPlayback.

📄

Professional PDF Export

Domain-specific PDFs via jsPDF with Hindi branding: Sanjeevani (Medical), Kavach (Security), Nyaaya (Legal), Prahari (Incident). Signature blocks, metadata grids, and colour-coded sections.

💻

Desktop App (Electron)

Cross-platform desktop via Electron 35 with GPU acceleration, WebAuthn support, and native packaging for Windows, macOS, and Linux.

🧠

Multi-Model Selection

Choose from 3 LLM tiers: Gemma 3 1B (~810 MB), Qwen2.5 0.5B (~400 MB), SmolLM2 135M (~100 MB). Match AI quality to your device’s capability.

♿

WCAG 2.2 Accessible

172+ ARIA attributes, full keyboard navigation, screen-reader support, reduced-motion and forced-colors media queries.

▸ TECH STACK

Component	Technology	Purpose
Framework	React 19 + TypeScript 5.9 (strict)	UI rendering and state management
Build Tool	Vite 7 + ESLint 10	Development server, bundling, code quality
AI SDK (core)	@runanywhere/web v0.1.0-beta.9	ModelManager, EventBus, OPFSStorage, VoicePipeline, VoiceAgent, SDKLogger
AI SDK (LLM)	@runanywhere/web-llamacpp	Streaming LLM, StructuredOutput, ToolCalling, Embeddings
AI SDK (Audio)	@runanywhere/web-onnx	STT (Whisper), TTS (Piper), VAD (Silero), AudioCapture, AudioPlayback
LLM Models	Gemma 3 1B / Qwen2.5 0.5B / SmolLM2 135M	3-tier model selection for all device capabilities
STT	Whisper Tiny English (ONNX) + Web Speech API	On-device speech-to-text with browser fallback
VAD	Silero VAD (ONNX, ~2.3 MB)	Real-time voice activity detection
TTS	Piper (Lessac, ONNX, ~64 MB)	On-device text-to-speech for voice feedback
Encryption	AES-256-GCM + HKDF + PBKDF2 (600K iter)	Per-case encryption via WebCrypto API
Auth	WebAuthn + PRF extension	Biometric auth (Windows Hello / Touch ID / Face ID)
PDF Export	jsPDF 4.2	Domain PDFs: Sanjeevani, Kavach, Nyaaya, Prahari
Desktop	Electron 35 + electron-builder	Cross-platform desktop with GPU acceleration
Fallback	Keyword Extraction Engine	Instant regex-based intelligence extraction
Offline	vite-plugin-pwa + Workbox	Full offline PWA with service worker caching
Testing	Vitest 4 + Testing Library	231 tests across 18 files

▸ ACCESS POINTS

🌐 Live Application 📂 GitHub Repository 🛡 Privacy Proof 🏗 Architecture 🇮🇳 India 2047 Vision

SHADOW NOTES