AI · Autonomous agent platform

Cortex Automator (Jriza)

An autonomous automation agent that drives real headless browsers, executes code in ephemeral Ubuntu sandboxes, and verifies its own output before it answers.

Live — web chat + public streaming API

jriza.in ↗

Overview

Cortex Automator is an autonomous AI automation agent that operates real headless browsers, executes code in ephemeral Ubuntu sandboxes, calls live APIs, and verifies its own output before responding. Where a conversational LLM describes what it would do, Cortex actually does it — running nmap scans, driving Chromium like a human, reverse-engineering HTTP/2 traffic, and shipping working artifacts back to the user.

The gap it closes: general-purpose chatbots break the moment a task needs side effects. They cannot open a browser and read the DevTools network tab, run nmap, sqlmap, ffuf or nuclei and interpret the output, clear CAPTCHAs and bot-detection surfaces such as Cloudflare, DataDome and PerimeterX, keep a ten-minute scan alive past a single request, or check an answer against reality before sending it.

The product ships as a web chat at jriza.in and a public streaming HTTP API, aimed at security researchers, automation engineers and technical teams who need an agent that executes and verifies rather than hallucinates.

Karnveer Singh led the platform build, the agent tooling and execution loop, and the SEO / AEO / GEO architecture for the public site.

Architecture

A task enters through a surface, is decomposed by the execution loop, and only leaves the system after a real browser, sandbox or API call has confirmed it.

  1. 01

    Surfaces

    Where a task enters the system.

    Web chat — jriza.in

    Streaming chat UI with live job strip

    Public HTTP API

    Same execution loop, streamed to external clients

  2. 02

    Execution loop

    Plan → act → observe → verify. Nothing is answered until a tool confirms it.

    Planner

    Decomposes the task, picks tools, retries and falls back

    Skill router

    30+ expert personas — Kali Operator, SQLMap, Burp, Web Scraper

    Verifier

    Re-checks output against observed reality before replying

  3. 03

    Execution surfaces

    The three places real side effects happen.

    Browser cluster

    Stealth headless Chromium via Browserless — human typing cadence, mouse jitter, auto-scroll, proxy rotation, CapMonster for reCAPTCHA v2/v3, hCaptcha and Turnstile

    E2B sandbox (per chat)

    Isolated Ubuntu box preloaded with nmap, sqlmap, metasploit, impacket, hashcat, python, node, git — destroyed after 5 min idle

    Live API calls

    HTTP/2 traffic capture, DNS recon and third-party APIs called directly

  4. 04

    Job runner

    Work that outlives the request.

    Detached runner

    nohup + disown wrapper; 10+ minute scans survive tab closes and new turns

    Status poller

    3-second polling drives elapsed time and the kill switch

  5. 05

    Streaming + artifacts

    What comes back to the user.

    Streaming terminal

    Token-by-token stdout in green, stderr in red, live exit codes and timers

    Artifact pipeline

    /home/user/uploads/ for drag-drop input, /home/user/out/ auto-detected and returned as signed download URLs, history persisted

Cortex Automator architecture — surfaces feed the execution loop, which drives the browser cluster, per-chat E2B sandbox and live APIs; long work is handed to the detached job runner, and results return as streamed terminal output plus signed artifact downloads.

Scope

  • Agent product and interaction design
  • Execution loop: plan, act, observe, verify
  • Headless browser cluster with stealth and CAPTCHA handling
  • Per-chat ephemeral Ubuntu sandboxes with a Kali toolkit
  • Long-running detached job runner and live job UI
  • Streaming terminal transport and artifact pipeline
  • Skills library of 30+ expert personas
  • Public streaming HTTP API
  • SEO / AEO / GEO architecture for jriza.in

Tech stack

Frontend

  • React
  • TypeScript
  • Tailwind CSS
  • Streaming chat UI
  • Live terminal renderer

Agent runtime

  • Tool-calling orchestration
  • Plan / act / observe / verify loop
  • Retry and fallback rules
  • Skill persona router

Browser layer

  • Headless Chromium
  • Browserless stealth profile
  • CapMonster (reCAPTCHA v2/v3, hCaptcha, Turnstile)
  • Proxy rotation

Sandbox layer

  • E2B ephemeral Ubuntu
  • nmap, sqlmap, metasploit, impacket, hashcat
  • python, node, git, curl
  • nohup + disown job runner

Backend

  • Node server functions
  • Job queue with status polling
  • Signed artifact URLs
  • Structured logging

Growth

  • JSON-LD structured data
  • Answer-engine content blocks
  • Core Web Vitals tuning

Deliverables

  • Real headless browser cluster — human-grade Chromium automation with human-like typing cadence, mouse jitter, auto-scroll, and CAPTCHA detection and solving
  • Per-chat ephemeral sandbox — every conversation spawns an isolated Ubuntu box preloaded with the Kali toolkit, auto-destroyed after 5 minutes of inactivity
  • Long-running background jobs — a detached runner lets scans and builds outlive the request, with a live strip showing elapsed time, 3-second status polling and a kill switch
  • Streaming terminal output — results stream token-by-token, stdout in green, stderr in red, exit codes and timers updating live
  • Sandbox file I/O — drag-drop upload into /home/user/uploads/, artifacts written to /home/user/out/ returned as signed download URLs, history persisted across sessions
  • Skills library of 30+ expert personas including Kali Operator, Red Team C2, Nmap, Burp, SQLMap, Metasploit, Impacket, Hashcat, Web Scraper and API Reverse Engineer
  • Public streaming HTTP API exposing the same execution loop to external clients
  • Marketing site for jriza.in with SEO, AEO and GEO architecture, schema markup and answer blocks

Outcomes

Surfaces
Web chat at jriza.in + public streaming HTTP API
Sandbox lifecycle
One isolated Ubuntu box per chat, destroyed after 5 min idle
Job durability
10+ minute jobs survive tab closes and new turns
Skills library
30+ expert personas

Frequently asked questions

What makes Cortex Automator different from a chatbot?
It executes real actions — driving a stealth headless Chromium, running commands in a live Ubuntu sandbox, calling APIs — then verifies the result before answering, instead of generating plausible text about what it would do.
What is in the per-chat sandbox?
An isolated Ubuntu box preloaded with the Kali toolkit — nmap, sqlmap, metasploit, impacket, hashcat — plus curl, python, node and git. It is created per conversation and destroyed after 5 minutes of inactivity.
How do long scans survive a closed tab?
A detached job runner wraps commands with nohup and disown so jobs outlive the request. The UI polls status every 3 seconds and shows elapsed time with a kill switch.
How does it handle CAPTCHAs and bot detection?
The browser layer uses a stealth Chromium profile with human-like typing cadence, mouse jitter and auto-scroll, proxy rotation, and CapMonster solving for reCAPTCHA v2/v3, hCaptcha and Turnstile against surfaces like Cloudflare, DataDome and PerimeterX.
Who built Cortex Automator?
Karnveer Singh built the platform, the agent execution loop and tooling, and the SEO, AEO and GEO architecture behind jriza.in.