Project : LeadRadar

Most lead tools ask you to describe your customer in a form. LeadRadar asks you to write a brief: what this instance hunts, where its bar sits, which sources count, how far back to look.

From that one brief it gathers signals from the web and from structured people data, scores and deduplicates what it finds, and drafts the outreach. A person approves before anything leaves. Every pass writes a short briefing, so you can read what it did rather than trust that it did something.

One organisation runs many instances. The brief is the whole configuration.

The retrieval design in this site’s chatbot is lifted from it.

What it is made of

A human-approved B2B lead-generation agent: hybrid retrieval over web signals and structured people data, orchestrated as a state machine, with a person approving before anything leaves. Built during the Turing College AI Engineering programme.

Its retrieval module is the one I ported into this site: BM25 keyword search and dense vector search, merged with reciprocal rank fusion. The algorithm came across; the Python infrastructure did not, because this site runs on Cloudflare Workers.

Why it exists

A friend who sells flowers asked me how to find customers. Then people who had built small tools asked the same thing about small and medium companies.

The question underneath was always the same: every channel that could find the right customer already exists, and nobody has time to watch all of them.

So give them to an agent. Let it search every day, watching the news that moves a market, new players arriving, technology shifting, and bring back leads.

I built it in 2026, as my third-sprint project at Turing College.