Privacy-First Browser Agent · Local Redaction Engine

AI that can see your browser. Without seeing your secrets.

A privacy-first browser agent that understands your screen, protects sensitive data locally, and takes action for you.

Download ExtensionChrome Extension · ZIPChrome Extension · ZIP

See How It Works

Try Gov Portal DemoInteractive Form Validation

Sensitive data is detected and redacted on-device before network transmission.

https://secure.bank.app/checkout

TLS 1.3 · SECURE

LOCAL SANDBOX: ON

PORTAL: FINSECURE CHECKOUT & KYC

Authenticated Account Tier 1 Verified

Alex Turner

alex.turner@enterprise.io

Account Email

Master Password

Payment Card (Luhn Checked)

Token Vault: 3 Keys Ephemeral In-Memory

SCANNING DOM & VISUAL BUFFER

Telemetry Status:CAPTURE · FRAMEBUFFER_IN

Zero Raw Exfiltration Guarantee

Your data stays yours.

Zero raw screen exfiltration. Cryptographic local isolation before any packet leaves your machine.

AIR-GAP DATA FLOW PIPELINESTRICT BOUNDARY

RAW DATAScreen Pixels & DOM

LOCAL PROCESSINGWebGPU & WASM Models

SANITIZED DATASafe Token Schema

AIRemote VLM Reasoning

See locally

Lightweight computer vision models run directly inside the browser.

WebGPU & WASM pipelines process viewport frames in client memory.

Redact locally

Sensitive information is detected and masked before network transmission.

Faces, cards, passwords, and QR codes are blacked out/blurred with zero roundtrip.

Reason remotely

Only sanitized context is sent to the AI model for complex reasoning.

The cloud VLM sees layout landmarks and surrogate tokens, never plaintext.

End-to-End Privacy Lifecycle

How netrAI Works

A seamless 5-step pipeline engineered to guarantee high-accuracy AI reasoning without ever exposing private user credentials.

01LOCAL CLIENT

Capture

Screen & DOM Ingestion

The extension observes the current browser screen and DOM.

Zero-latency canvas framebuffer capture & selective DOM tree observer.

02LOCAL WASM/WEBGPU

Detect

On-Device Neural Scan

Local vision models and privacy rules identify sensitive information.

BlazeFace, RapidOCR, and Luhn/Aadhaar/PAN regex rule engines run concurrently.

03AIR-GAP VAULT

Redact

Pixel Masking & Vaulting

Sensitive regions are masked and replaced with safe tokens.

Canvas pixel obscuration & surrogate token registration (#TOK_CRED_8891).

04CLOUD SANITIZED

Reason

Sanitized Ingress VLM

Sanitized visual and DOM context is sent to the AI reasoning layer.

Multimodal VLM receives anonymized landmarks with zero plaintext PII.

05LOCAL DOM ENGINE

Act

Synthetic Local Execution

The AI returns structured commands (click, fill, scroll) that the browser executes locally.

Synthesized JSON actions dispatched locally with token vault re-injection.

ACTIVE STAGE: STEP 01 — CAPTURE

The extension observes the current browser screen and DOM.

On-Device Neural Stack

The browser does the sensitive work.

Embedded WebGPU & WebAssembly neural engines executing real-time computer vision entirely on-device with zero raw cloud roundtrips.

INTERACTIVE PIPELINETAP A STAGE TO INSPECT

→→→→→→

~300MBONNX Web Runtime

CLIP ViT-B32

Semantic Vision & UI Layout Parsing

Generates rich visual embeddings to map complex interactive elements and web application layout structures without sending raw framebuffers off-device.

On-Device ExecutionAir-Gapped

~1.5MBWASM / WebGPU

BlazeFace

Ultra-Fast Facial Landmark Detector

Sub-millisecond facial bounding box locator optimized for mobile GPU architectures, ensuring KYC selfies and user profile avatars are detected instantaneously.

On-Device ExecutionAir-Gapped

~4.2MBONNX WASM

RapidOCR

Client-Side Optical Character Recognition

Extracts alphanumeric strings from canvas captures, buttons, labels, and receipt screenshots purely on the CPU/GPU thread inside the extension sandbox.

On-Device ExecutionAir-Gapped

~45KBPure JS / WASM

jsQR

Zero-Dependency QR & Barcode Locator

Real-time binarization and parsing of authentication 2FA QR codes, payment UPI QR matrices, and shipping barcodes directly in browser memory.

On-Device ExecutionAir-Gapped

DeterministicNative V8 Heuristics

Regex & Checksum Engine

Luhn Algorithm, Aadhaar, PAN & SSN Verifiers

High-speed algorithmic checks validating payment card mod-10 Luhn checksums, Indian Aadhaar Verhoeff verification, PAN alphanumeric tokens, and US SSNs.

On-Device ExecutionAir-Gapped

Direct HardwareChromium GPU Pipeline

WebGPU

Hardware-Accelerated Tensor Computation

Harnesses client GPU compute shaders for parallel neural inference, keeping execution responsive and avoiding CPU thread blocking during heavy web browsing.

On-Device ExecutionAir-Gapped

Memory-SafeSandboxed Bytecode

WebAssembly (WASM)

High-Throughput Pixel Sandbox

Compiles C++ and Rust computer vision algorithms to near-native speed inside an isolated, memory-safe execution sandbox with strict zero-system-access guarantees.

On-Device ExecutionAir-Gapped

Zero-Trust Cybersecurity Perimeter

Zero-Trust Privacy Architecture

An immutable cryptographic barrier partitions local client sandbox memory from remote cloud inference. Raw secrets cannot cross the perimeter.

LOCAL / TRUSTED ZONE100% ON-DEVICE

Chromium Extension Sandboxed Memory & WebGPU

01. Screen captureNative Canvas API

Extracts ephemeral viewport framebuffer and DOM node hierarchy in client RAM.

02. Local visionBlazeFace & OCR WASM

Scans screen pixels for faces, avatars, and alphanumeric text with zero cloud roundtrips.

03. PII detectionLuhn / Aadhaar / PAN Rules

Verifies credit card checksums, tax IDs, medical records, and credential fields.

04. Redaction engineGaussian & DOM Masking

Blurs visual bounding boxes and replaces DOM strings with surrogate token IDs.

05. Token vaultEphemeral In-Memory RAM

Isolated ephemeral browser-memory dictionary mapping real data ↔ surrogate tokens. Never written to disk or sent across networks.

06. Action executorPure Local DOM Dispatch

Executes verified structured action JSON directly against page DOM elements, re-injecting vaulted tokens at submit time.

RAW PERSONAL DATA ✕ NEVER SENT

CLOUD / SANITIZED ZONEREMOTE VLM

Multimodal Large Vision-Language Model

01. Sanitized Image & Sanitized DOMSanitized Payload

Multimodal image containing only blurred blackouts, with stripped DOM layout trees and surrogate token IDs.

02. VLM ReasoningChain-of-Thought

Performs visual workflow reasoning: analyzes button coordinates, form structure, and navigation intent.

03. Action command generationStrict JSON Action Schema

Compiles reasoning into safe JSON commands (e.g. { action: "click", selector: "#submit" }) rather than executable JavaScript.

EGRESS AUDIT REPORT

• Plaintext Passwords Transmitted: 0
• Unredacted Face Pixels: 0
• Credit Card PAN Exfiltration: 0%

SANITIZED CONTEXT ✓ SAFE TO PROCESS

Dual-Path Decision Engine

Intelligent Dual-Path Routing

Not every web action requires a remote supercomputer. netrAI routes simple interactions locally and reserves cloud VLMs for high-complexity visual reasoning.

CORE EVALUATION LOGIC

Can this task be completed locally?

YES BRANCH — ZERO NETWORK CALL

Local rules / local model

DOM heuristics, deterministic selectors, and lightweight client models execute directly in browser memory.

NO BRANCH — SANITIZED CLOUD INGRESS

Sanitized context → Server VLM

Multimodal visual reasoning dispatched exclusively with masked screen bounding boxes and tokenized strings.

Select Interactive Task Scenario:

EVALUATED ROUTING DECISION

YES — Complete Locally

Target Engine:Local rules / local model (DOM Heuristics)

Task Command Prompt:

"Scroll to pricing table and click monthly billing toggle"

Network & Privacy Guarantee:

0 KB (Offline / Air-Gapped)

Routing Rationale: Exact CSS selector and scroll destination detected directly in local DOM. No visual reasoning required.

Defense-in-Depth Engineering

Privacy isn't a setting. It's the architecture.

Defense-in-depth engineering designed so that even a fully compromised server cannot view your private information.

CLIENT LOCAL MEMORY

On-device detection

Inspects visual pixels and raw DOM node contents entirely in local memory before any HTTP payload or network packet is constructed.

WebGPU and WASM neural inference without cloud dependencies.

AIR-GAP REDACTION

Redaction before network

Permanent client-side graphical and string redaction occurs before network sockets or WebSockets open. Visual bounding boxes are pixel-blurred.

Mathematical zero-knowledge egress for sensitive pixels.

IN-MEMORY RAM VAULT

Local token vault

Raw passwords, Aadhaar, PAN, and card numbers exist exclusively inside ephemeral in-memory token vaults with zero persistence to disk or remote servers.

Ephemeral key-value RAM mapping cleared on tab close.

SURROGATE TOKEN INGRESS

Sanitized reasoning

Server VLMs operate purely on structural tokens ([TOKEN_CARD_1], [TOKEN_NAME]), making credential extraction mathematically impossible.

Cloud VLM cannot leak what it never receives.

ANTI-RCE SCHEMA

Structured actions

Remote AI produces strict JSON action schema ({ action: "click", selector: "..." }) rather than executable scripts, preventing remote code execution (RCE).

Eliminates arbitrary script injection and remote code execution (RCE) vulnerabilities.

Architecture Performance

Engineered for Uncompromising Speed

By shifting visual intelligence and privacy redaction directly onto the user's machine, netrAI eliminates the massive network overhead of streaming raw screen video to the cloud.

ZERO ROUNDTRIP

Local inference

Sub-frame computer vision and element localization executing directly via client GPU shaders without network latency.

WebGPU compute shaders execute bounding box detection concurrently with page render cycles.

ANONYMIZED SCHEMA

Low network payload

Downscaled, optimized, and redacted screenshot payloads reduce network bandwidth by orders of magnitude compared to raw visual streaming.

Transmits compact structured coordinate graphs and masked fragments rather than continuous uncompressed video streams.

CLIENT MEMORY MASK

On-device redaction

Instantaneous tokenization and pixel masking with zero roundtrip cloud dependency. PII is obscured before sockets connect.

Canvas shader operations obscure pixel buffers synchronously in browser memory before JSON serialization.

NATIVE DOM EVENTS

Browser-native execution

Direct execution in Chromium extension service worker with native DOM event dispatching for instant UI responses.

Synthetic click and fill actions invoke native browser event dispatchers without external remote desktop layers.

QUANTIZED WASM

Lightweight models

Quantized neural networks engineered and pruned specifically for modern web browser runtime constraints and client hardware.

ONNX Runtime Web with SIMD acceleration ensures minimal memory footprint and zero browser freezing.

Interactive Live Simulation

See netrAI in Action

Experience how the visual privacy agent safely completes multi-step tasks in real time across sensitive web applications.

Scenario:

https://finsecure.pay/checkout

PAYMENT GATEWAY

Fintech Checkout Gateway

Simulated Web Viewport

Air-Gap Active

Cardholder Name

Aadhaar / SSN

Card Number

Target: #submit-btn

netrAI Agent Telemetry Log

STAGE 1/7

Current Lifecycle State:STATUS: OK

Screen captured

Extension ingests viewport canvas buffer and DOM elements into local RAM.

Active Telemetry Stream:

> [DOM_OBSERVER] Ingested 142 DOM nodes, 1920x1080 canvas frame loaded into WASM memory.

Dispatched Action Intent:

CAPTURE_FRAMEBUFFER_OKJSON SCHEMA

Local Token Vault Status:0 Keys (Pre-Scan)

Open-Source Local Privacy Architecture

Give AI your browser. Not your private data.

Download the privacy-preserving browser agent and experience AI assistance without handing over your sensitive screen data.

Download ExtensionChrome Extension · ZIPChrome Extension · ZIP

View Architecture

Zero Cloud TelemetryManifest V3 CompliantEphemeral Token Vault