joehahn/prospect-scanner

**I built an AI system that finds consulting prospects in public sources, decides
which are worth my time, drafts the first note, revises it with me until it sounds
like me, and grades its own work.** I'm a one-person AI practice, this is how I find
my own clients, and I use it nearly every day.
Finding clients is mostly research. Which firms just did something that gives me a
reason to write? Is this one worth an hour? Who there can actually say yes? This
does that research from public evidence, so hours of reading become minutes of
deciding. Each morning I get a short list of people and, when I ask, a draft note to
each. I edit and send every note myself, by email or LinkedIn.
A working morning. The firms and prospects are fictional; the LinkedIn profile is my own.
**Author:** Joseph M. Hahn, Ph.D., independent AI and machine learning consultant
jmh-datasciences.com · LinkedIn · joe.hahn@jmh-datasciences.com
**Built end-to-end with Claude Code.** · **License:** PolyForm Noncommercial
---
What it does
- **Finds people with a dated reason to talk.** A new AI leader named, a vendor
chosen, a fund raised, a conference talk about a problem worth solving. Not lists
of names.
- **Vets each firm from its own website.** What it does, how big it is, who decides,
and for an investment firm, whose boards each partner sits on.
- **Rates every prospect with an AI judge that learns from me.** There are no scoring
rules. It learns from my calls on similar people, Write first or Wouldn't, and gets
better with each one.
- **Drafts the note in my voice,** from notes I actually sent, then checks every
factual claim against its source before I see it.
- **Measures itself.** Every model call is costed, every claim is sourced, and the AI
judge is scored against a plain formula run beside it, using my own write and
don't-write calls as the answer key (how).
Blue: a single LLM call (Claude) for reading, judgment or writing, each with a structured output. Purple: an AI agent that uses tools. Grey: deterministic code, search, and the database every step reads and writes. Amber: you.
Where AI is used, and where it isn't
Each step gets the least powerful tool that does the job well. The strongest model,
Claude Opus, writes the notes, because a note is the one thing a stranger reads.
Reading hundreds of search results and web pages a day is high volume and needs less
judgment, so smaller, cheaper models do that. Anything that can be computed uses no
AI at all.
| Step and model | What it does, and why AI rather than code |
|---|---|
| **LLM reading** Claude Haiku, Sonnet | Turns Tavily search results and web pages into structured facts: keeps only real, dated events, and pulls out what a firm does, who decides, and whose boards a partner sits on. Finding a firm's own website uses Claude with Anthropic's web search tool. Pages vary endlessly, and a fixed schema keeps the output checkable. |
| **LLM screening** Claude Haiku | Asks the judge's question once, with the same prompt and evidence, on the cheap model, so the judge's three runs go only to people worth them. I measure how well the screen predicts the judge, and set its pass mark from that. |
| **LLM judgment** Claude Sonnet | Rates each prospect 1-5 from my own Write first / Wouldn't calls, and works out what the person is after and which offer fits. My taste is learned from examples, not written down as rules. |
| **LLM writing** Claude Opus | Works out the note's storyline, drafts it, and revises it when I ask. A separate call then checks every claim against its source before the note is saved. |