I built my own company-discovery agent: Exa semantic search + a Claude scoring pass

A little agent that hunts for companies I should keep an eye on for my job search — before they even post an opening, so I already know about them the moment they start hiring. Here’s how it works: it searches the web using Exa, a search engine that understands meaning rather than just keywords, so I can describe the kind of company I’m looking for in plain sentences instead of guessing the right search terms. It checks each new company against a list of ones I already know about, so it never wastes time re-checking the same company twice. Then Claude (Anthropic’s AI model) reads each new company’s website and scores it from 0 to 10 based on how well it fits what I’m looking for. Anything that scores high enough gets sent straight to my phone via Telegram. The whole thing runs automatically on GitHub’s free infrastructure — no server to maintain, no database. It just saves its results as simple files back in the same code repository, so that becomes its “memory” between runs.

I’ll run it live on screen and show three things most people never see: (1) the raw data that comes back from the search engine — exactly what it finds and how I pull the useful text out of it; (2) the messy backup logic I had to write for when the AI doesn’t format its answer properly — because it doesn’t always follow instructions perfectly, so I had to teach the code to clean up after it and recover instead of crashing; and (3) the actual message landing on my phone at the end. I’ll also show the single most important decision in the whole system: the moment the AI has to figure out whether a company actually builds a product, or just sounds like it does while really being a consulting agency.