The Problem: AI Can Talk, But It Can’t "Do"
We’ve all seen LLMs write incredible code and prose. But when you ask an AI to "Go to Stripe, find my last three invoices, and upload them to Slack," things get messy. Traditional RPA is brittle, and most "AI Agents" get lost the moment a div class changes.
That’s why I’m building Project Ghost.
What is Project Ghost?
Project Ghost isn't just another scraper; it’s a Universal Action Layer. It’s designed to bridge the gap between high-level LLM intent and low-level browser execution. Instead of hard-coded selectors, it uses AI-powered browser automation to "see" and interact with the web just like a human would.
The Tech Stack 🛠️
To keep things fast and scalable as a solo dev, I’ve leaned into a modern, serverless-first stack:
- Frontend/API: React & Vercel
- Backend/Database: Supabase (for Auth and lightning-fast data handling)
- Language: Python (the heavy lifting for automation logic)
The Core Challenge: Solving the "Brittle Selector" Problem
The biggest hurdle in browser automation is that websites change. Project Ghost handles this by:
- DOM Semantic Analysis: Stripping the junk and sending only the relevant accessibility tree to the LLM.
- Self-Healing Actions: If an element isn't where it’s expected to be, the agent re-evaluates the page state in real-time. Link: https://project-ghost-lilac.vercel.app/
Top comments (0)