While I sleep, an AI checks 3,900 Japanese stocks (Part 1)

This is the English edition of a five-part series originally published in Japanese, documenting how an individual investor — not a programmer — built an automated research system for the Japanese stock market together with AI agents. Failures included.

Here is a strange fact about the Japanese stock market: it has more than 3,900 listed companies, and 62% of them are covered by zero analysts. Not "under-covered". Zero. The average small cap here is followed by 0.9 analysts, and English disclosure only became mandatory in 2025 — and only for the largest ~1,600 companies on the Prime market. Everything else is, for most of the world, a blind spot.

I live inside that blind spot. And every night, while I sleep, all 3,900+ of those companies get checked against the same criteria.

The one doing the checking is not me. It's a "fixed-point observation system" I built together with AI agents. Every night it ingests prices, financials and regulatory disclosures, recalculates scores, and if something changed, a report is waiting in my Slack the next morning. It schedules its own daily posts to my Japanese blog, and posts to X five times a day (one of them in English) — by itself.

Written like that, it sounds smooth. It was not. This series is the record of a non-programmer building "personal investment infrastructure" hand-in-hand with AI — including the times the data betrayed me, and the time a fake market crash nearly went out to paying readers.

Why I built it: I got tired of reading

When you start investing, you drown in opinions. Social media is full of "this one's next". The news delivers a different mood every day. The more you read, the less you know what's real and what's just convenient for the person saying it.

At some point I realized: what I actually wanted wasn't anyone's opinion. It was the same criteria, applied to every company, every day.

No human can do that. There are legendary investors in Japan who read the entire Shikiho (the quarterly handbook that profiles every listed company) cover to cover — but nobody does it daily. "Same criteria, every day, every stock" is a job only a machine can do.

What the system does

• Every night, it ingests prices, financials and disclosures for all 3,900+ listed companies
• It computes growth, profitability, balance-sheet safety and valuation, and sorts them into plain tiers (high / mid / low)
• It splits every declining stock into two piles, daily: healthy companies being sold off ("dip candidates") versus declines with deteriorating fundamentals ("dangerous drops")
• Every week and month, it scores its own past calls — what actually happened to the candidates it flagged
• Every morning at 9:45, it runs a 25-point self-inspection and reports anomalies to my Slack
• 55 automated jobs run on daily, weekly and monthly schedules
• Publishing is automated too — daily reports and five X posts a day

One number to convey the flavor: the system has over 1,800 regression tests — automated checks that previously-fixed bugs haven't come back, all of which run every time we touch the code. Excessive for a personal project? Maybe. But this thing handles numbers that people might act on with real money. I don't compromise there.

How it was built: I didn't write most of the code

An honest confession: the majority of this system's code was not written by me. It was written by an AI agent.

My job is three things: say what I want, decide at the forks in the road, and doubt what comes out until it's verified against real data.

There's a twist, though. This project uses a second AI whose only job is to hunt for bugs in the first AI's code. One writes, the other reviews — the same adversarial code review human teams have always done, except both reviewers are machines. It has caught real, product-breaking mistakes before they shipped, repeatedly.

"You trust AI with this?" — the correct answer is no. That's exactly why there are 1,800+ tests, a morning self-inspection, and adversarial review. Trust is something you build out of mechanisms, not vibes. That's the biggest lesson so far.

The one thing I care about most: the scorecard

The system has a feature we call answer-checking. Every candidate it ever flagged is mechanically re-scored — weekly and monthly, wins and losses, published.

Showing only the winners would be easy. But do that and you fool yourself before you fool any reader. Recording every miss isn't just honesty toward others; it's the only way to actually improve your own judgment.

So no — this system does not "predict the next winner". It's a machine that helps you practice following the numbers yourself, and never lets your past calls quietly disappear. That distinction will come up again and again in this series.

There are plenty of failures to tell

A preview of what's coming:

• Price ingestion once silently stopped for four and a half months — with every job reporting "success" the whole time
• Corrupted financial data once gave a major company an impossible EPS — which led to the "corruption firewall" that keeps broken numbers out of anything readers pay for
• Just recently, a wave of stock splits created data seams that made blue chips look like they'd crashed -91% in a week — a fake crash caught 90 minutes before publication

Data betrays you. Every betrayal becomes another detector and another firewall. Building a system, it turns out, is the craft of converting failure into structure.

Next in Part 2: what it's actually like to delegate development to AI agents — how to give instructions, and how to argue with them.

Tsukiyo Research publishes a weekly systematic report on Japan's uncovered small caps — every issue opens with the scorecard of our past calls. If this series interests you, subscribe here.

This series is a personal build-and-operate log. Nothing here is investment advice or a recommendation to buy or sell any security.