Skip to the work

ClientMine: how it is built

A weekly board of companies ready to buy, built by AI agents and approved by a person.

clientmine.aiClientMine, opens in a new tab

ClientMine is a web app that tells a firm, every week, which companies are ready to buy. It checks two things together: signals of buying readiness, like hiring abroad, new funding or SEC filings, and fit, meaning the right type of company with a clear decision maker. Each company comes with a tailored approach, and a person approves the board before the firm sees it. Its first paying client is an immigration law firm.

Tools. Claude Code agents, Python scraping, Firecrawl, SEC EDGAR API, PredictLeads API, Supabase, Next.js, Vercel.

The technical write-up

Run a weekly scan. Watch which steps are AI agents and which are plain code.Sample

  1. Find companies with a buying triggerAIScout agent
  2. Read each company’s website onceCodeCode
  3. Check the gates, score the fitAIScorer agent
  4. Research the best, up to 6 a weekAIEnricher agent
  5. Write an approach for eachAIEnricher agent
  6. Final score and tierCodeCode, no AI
  7. You approve the boardYouA person
  8. Publish to the firm’s boardCodeCode
  9. Recheck last week’s companiesAIMonitor agent, alongside

This week’s scan hasn’t run yet.

Last week’s board is still live for the firm. Press Run it to watch the agents build the next one.

  1. Larkspur Biotech

    Raised a Series B in Germany

    Site readTier A

    Head of People. Email found.

    First moveCongratulate her on the raise. Ask how the German team will move over.

  2. Tidewell Analytics

    Opening a U.S. office

    Site readTier A

    COO. Email found.

    First moveAsk who handles visas for the U.S. launch team.

  3. Oakmere Foods

    Local bakery, no hiring

    Site read

    Too small
  4. Brightline Freight

    Hiring 12 engineers abroad

    Site readTier B

    VP Engineering. On LinkedIn.

    First moveAsk about timing for the 12 hires abroad.

  5. Quillon Software

    New VP of People

    Site readTier B

  6. Marrow & Pine

    A law firm itself

    Site read

    Wrong industry
  7. Halcyon Labs

    Expanding to Texas

    Site readTier B

  8. Vessel Robotics

    No named leader to reach

    Site read

    No way in

Hot ListNew this week

Acting now

    Worth building

      How it was built

      What the firm got

      • A board of companies to call, each with why now, a fit score, who to contact and a first move.
      • Research spent where it counts. Cheap checks drop poor fits first, and deep research goes to at most 6 companies a week.
      • Nothing goes live without a person. The new board waits for approval, and a failed scan never replaces last week’s board.

      How I built it

      1. Split judgment from mechanics

        AI agents do the reading and judging: finding companies, scoring fit, researching contacts and watching for changes. Plain Python runs the order, the rules, the final score and the saving.

      2. Hand work over in files

        Each agent writes its results to a file in that week’s folder. The next step reads it, and a checker rejects any file that breaks the agreed shape.

      3. Put a person at the gate

        The finished board waits for approval, tied to that exact board. The database write refuses to run without it.

      4. Ship it as a product

        A Next.js app on Vercel, Supabase for each firm’s data and board, and PredictLeads for hiring and funding signals. The first paying firm joined in July 2026.

      5. Test every rule

        About 170 automated test files cover the plain-code steps, the file checks, the gates and the database write. A failed scan never replaces last week’s board.

      The agents

      • Scout finds companies showing a buying trigger.
      • Scorer checks the gates and scores fit.
      • Enricher researches the best companies, finds who to contact and writes an approach for each: a first move, a draft first message and questions for the call.
      • Monitor rechecks earlier companies for changes, alongside the rest.

      Each agent is a Claude Code agent following a written procedure. Everything between them is plain Python.

      Back to the home page

      How it is built, for the technical reader.

      One product, end to end: what I did, how it is put together, what it costs to run well, and where it falls short.

      Tools. Claude Code agents, Python scraping, Firecrawl, SEC EDGAR API, PredictLeads API, Supabase, Next.js, Vercel, React, Tavily, Apollo.

      My part
      I designed and built the product: the weekly scan, the agents’ written procedures, the database and the web app the firm uses. My partner sells it and runs the client relationship.
      The design
      AI agents do the reading and judging. Plain Python runs everything else: the order of steps, the gates, the final score and the save. Each agent writes its results to a file, and a checker rejects any file that breaks the agreed shape before the next step reads it.
      One trade-off
      Deep research costs the most, so it is spent last and on at most 6 companies a week. Cheap checks drop poor fits first. The firm gets fewer companies, each researched properly, instead of a long list nobody can call.
      How it is tested
      About 170 automated test files cover the plain-code steps, the file checks, the gates and the database write. The approval gate is enforced in code: the write refuses to run without an approval tied to that exact board.
      Limits
      The weekly scan still runs with me supervising. Outside data services can fail or hit a usage limit mid-scan. When a step fails, nothing is published and last week’s board stays up.
      Status
      In paid use since July 2026 by an immigration law firm, with a new board each week.