Demo working preview

🧠 Browser Automation Agent — Small Steps, Big Promise

Ever wished you had an AI assistant that could browse the internet for you?

Meet our Browser Automation Agent — a proof-of-concept agent that mimics a human-like browsing experience. It's powered by LLMs and a headless browser, capable of navigating websites, logging in, extracting data, and even comparing prices... all hands-free.

⚠️ Not ready to conquer the Voeger Challenge just yet... but it's a promising start 😅

🔗 Project Links

⚙️ Tech Stack

Here's a peek into the technologies powering this browser automation agent:

🧠 AI Models

LLaMA 3.3 — Responsible for parsing user prompts and generating structured step-by-step plans
LangChain — Powers agent routing, planning, and memory around the LLM

🧰 Frameworks & SDKs

FastAPI — Backend framework orchestrating prompt handling, execution, and response streaming
Playwright — Headless browser automation for navigating, interacting, and scraping real web pages
Chromium — Browser engine used under the hood by Playwright
WebSockets — Enables real-time streaming of updates between the agent backend and frontend

💻 Code & UI

Vite — Lightning-fast frontend bundler powering the minimal dashboard
Tailwind CSS — Utility-first styling for rapid UI development
Zustand — Lightweight global state management for real-time updates

🚀 Together, these tools enable natural language → autonomous browser navigation → structured insights.

🎬 What Can It Do? (Tested & Demoed)

✅ Task 1: Google Flights 🛫

Prompt: “Search for Mumbai to Delhi flights”
Goal: Find and return the cheapest option available.
✔️ Successfully parses flight options from search results.

✅ Task 2: GitHub Login & Dashboard Extraction

Prompt: “Login to github.com with username/password and get dashboard data”
Goal: Perform authorization and fetch user-specific data.
✔️ Navigates login flow and scrapes dashboard contents post-auth.

✅ Task 3: TSLA Stock Check on Yahoo Finance 📉

Prompt: “Go to yahoo.finance.com, search TSLA, and fetch stock price report”
Goal: Retrieve Tesla's real-time stock price and summary.
✔️ Successfully extracts key financial metrics.

✅ Task 4: Amazon India Product Search 🛒

Prompt: “Search for 4K monitors on amazon.in and fetch results”
Goal: Return list of monitors with title, price, and rating.
✔️ Handles search input and scrapes top listings from Amazon.

✅ Task 5: Google > CrustData Website 🧭

Prompt: “Search CrustData on Google, visit site, and gather company info”
Goal: Navigate through Google, land on company site, and extract key data.
✔️ Completes multi-step navigation and info extraction.

🔧 Under the Hood

All tasks were prompted using plain English.
The agent uses a headless browser powered by automation scripts and enhanced with LLM-driven decision making.
Navigation, clicking, form-filling, and scraping are all handled autonomously.

Stay tuned — next up, we’ll break down the working internals, the stack, how it reacts to DOM changes, and how it could evolve into a full-blown agent for real-world tasks.

🛠 Setup Instructions

Want to try it yourself?

👉 Check out the GitHub repository for detailed setup & deployment steps

🧪 Under the Hood — Core Technicals

The agent is orchestrated using a FastAPI backend, which handles incoming natural language prompts and translates them into step-wise browsing plans using LangChain. These plans are then executed with Playwright, a powerful headless browser automation framework that enables the agent to navigate, interact with, and extract data from real-world websites just like a human would.

At the heart of the decision-making lies LLaMA 3.3, hosted on Groq, which provides ultra-fast reasoning to interpret user goals, generate structured actions (NAVIGATE, INTERACT, EXTRACT), and extract relevant information from raw page content. The browser interacts via actions like fill, click, or fill-click, based on contextual cues and objective-driven parsing. A streaming system continuously relays status updates back to the user, making the whole experience feel live and interactive.

From parsing flights to scraping stock reports, this agent shows what’s possible when LLMs meet browsers.