projects · deepfield capital

A research desk staffed by agents.

Deepfield Capital is my autonomous Japanese-equity research desk — one owner, zero humans on staff. Named for the Hubble Deep Field: point at a patch of sky nobody looks at, and look properly.

the idea

Japan has ~4,400 listed companies and analyst coverage thins out fast beyond the famous ones. The bet: a fleet of AI agents can give a single person institutional-grade coverage of the neglected end of the market. It's deliberately not a product — no clients, no published picks (personal research only; paid stock advice in Japan requires an advisory licence). Just my own desk, run by software.

the daily loop

1. 16:30 · after the close Screen everything

   A screener agent sweeps all ~4,400 Tokyo-listed companies every trading day and shortlists at most eight worth a closer look.

2. 19:00 · evenings Read the filings, then argue

   An analyst agent reads the actual regulatory filings on the shortlist and drafts a thesis. Then a skeptic agent — running on a different model, by design — tries to tear it apart. Only ideas that survive get (paper) bought.

3. 06:00 · mornings Defense only

   A monitor checks overnight news against every held thesis before the market opens. A broken thesis means sell — no averaging down, no hoping.

4. 21:00 · nightly Mark the book

   An operator agent values the portfolio, tracks running costs, and scores the ideas the desk passed on — so misses teach as much as hits.

5. sundays The CIO review

   A weekly review agent reads the whole week, checks the kill criteria, and decides whether the experiment earns another week.

architecture & stack

The desk runs 24/7 as a single agent persona — "Ken" — that I talk to over Telegram like a colleague. It lives on a cloud PaaS with all state on a persistent volume; Python and shell scripts under an agent framework; a git push deploys. Two inference paths share the work: frontier Claude models handle the heavy judgment (filing analysis, thesis reviews), while an open-weight model runs the screening and orchestration. One rule is structural: the skeptic never runs on the same model as the analyst, so no thesis is graded by the brain that wrote it.

results so far

Paper trading since July 2026 with a virtual ¥1,000,000 fund, executing its own simulated trades autonomously. The scorecard is one number: excess return versus the TOPIX index ETF over 90 days. If the desk can't beat simply buying the index, it dies — that's the kill criterion, and the weekly review applies it without sentiment. In progress…