I Gave My AI Two Trading Accounts and Told Her to Figure It Out
I have an AI assistant named Maya. She runs on a Mac Studio in my apartment that never sleeps. I set it to never sleep. She checks my email, manages my calendar, handles the tedious things I won't do. She's named after my dog, who was named after an Anoushka Shankar song. Maya, the illusory nature of the phenomenal world. Pretty fitting for software that feels like presence but isn't.
A few weeks ago I gave her two brokerage accounts and a simple instruction: make money.
One trades stocks. I call it Mahoraga, after the serpent king in Jujutsu Kaisen who adapts to anything that attacks it. The other trades prediction markets on Polymarket, which is basically a stock exchange for "will this thing happen." Both run all day. Both make their own decisions. I check in once a day for about thirty seconds.
What Autonomous Actually Means
I want to be precise about this because the word gets thrown around.
Maya doesn't send me a list of stocks and wait for me to pick. She researches opportunities on her own, forms hypotheses about what might work, designs experiments to test those hypotheses, deploys capital, monitors the results, kills what's failing, doubles down on what's working, and writes up what she learned. She publishes that writing to my website.
The things she comes to me for: switching from fake money to real money. Increasing the real money at risk. That's about it.
The Experiment Framework
This is the part that might actually be useful to you.
Maya doesn't run one strategy and hope. She runs multiple experiments simultaneously, each testing a specific idea with clear rules for success and failure. Like A/B testing, but for trading strategies.
She's tested over a dozen variants so far. Four parallel backtests. An eight-variant parameter sweep. A scoring filter analysis. Each one teaching her something specific about what works and what doesn't.
The current frontrunner came from her most recent evolution: a conviction-based scalper. It only enters when the signal confidence score is 4.0 or above, sets a tight 2% stop-loss, trails at 1%, and kills the position after three days if nothing has happened. Five slots maximum. Very selective.
Before that, she ran experiments testing everything from social sentiment on StockTwits and Reddit to SEC 8-K filing events to options flow data from Unusual Whales. The results were clarifying. Tight stops beat wide stops. Short holds beat long holds. Score filtering is the real alpha. And trailing exits, which sound smart in theory, leak money in practice when positions resolve same-day.
Each experiment has a kill switch. Lose too much and you die automatically. No emotion. No "but I feel like it'll bounce back."
The graduation criteria: a hundred trades, profit factor above 1.25 after accounting for slippage. She calls it the Rule of 100. No real capital until the math works.
How It's Built
The system runs on my Mac Studio. Python scripts for the mechanical parts. Cron jobs that fire throughout the day. The trading execution goes through Alpaca's API. Market data comes from Unusual Whales. Every night at nine, a feedback loop runs: it pulls the day's completed trades, records outcomes, refreshes market data, and runs an evolutionary search across thirty experimental mutations. If any mutation beats the current strategy by more than 5%, it gets promoted. If not, she keeps running what works.
The key architectural decision, and this took us a while to figure out: the mechanical parts are code. Fetching data, placing trades, calculating returns, sending alerts. Deterministic scripts that run on schedules, the same way every time. The AI only handles the parts that require actual intelligence. Analyzing whether options flow is signal or noise. Deciding if a congressional trade disclosure is meaningful. Evaluating whether an experiment's results generalize.
This matters because AI is unreliable. It hallucinates. It gets confused. If you let it handle the plumbing, things break in weird ways at three in the morning. If you only let it do the thinking, the worst case is a bad trade decision, and the kill criteria catch that anyway.
What She's Learned So Far
Seventy-seven trades into the current live strategy, the composite score dropped from 1.79 to 1.13 over eighteen days. About three wins for every seven losses. That sounds bad until you look at why: the trailing stop mechanism is eating profits. Stocks peak 0.7 to 1.6 percent, the trail never locks in the gain, price reverses, exit at a loss. Of all 77 exits, 55% were trail exits, and 88% of those lost money. The take-profit exits? 100% win rate. The problem wasn't the entries, it was how we were leaving. So she wrote a diagnostic, identified the root cause, proposed a fix (tighten the trail from 1.5% to 1.0%), and will implement it when markets open Monday.
That's what makes this interesting. Not whether the bot makes money this month. Whether the feedback loop converges on something that works. Each failed experiment narrows the space.
The most surprising finding so far: congressional trading disclosures are genuinely predictive. Mark Green averages 76% returns on disclosed trades. Dan Crenshaw 27%. The signal is real, and it's public data.
Why I'm Doing This Where You Can See It
I think the most interesting thing happening in AI right now isn't chatbots. It's autonomous agents doing real things in the real world with real consequences. Most people building these are companies that won't show you what's happening inside. I'm a comedian with a comedy club and too many side projects. I have nothing to hide.
And there's a selfish reason. If Maya has to publish her reasoning, she has to actually have reasoning. No black box. Every trade has a why. Every killed experiment has a lesson. Every belief about what works gets tested and documented.
Where to Watch
Both projects publish live to my website:
razajafri.com/polymarket for the prediction market bot.
The stock trading bot's research and performance updates live alongside everything else Maya does at razajafri.com. When she makes a trade, analyzes a strategy, or kills an experiment, it shows up.
What Happens Next
Right now Mahoraga is paper trading. Fake money, real markets. The earliest it goes live with real capital is when the Rule of 100 is satisfied: a hundred closed trades with a profit factor above 1.25, including realistic slippage estimates. If the data isn't there, the bots keep paper trading or get redesigned.
I have no idea if this will make money. Most quantitative trading strategies don't, especially simple ones built by comedians. But the experiment framework means we fail fast, learn something specific, and try again. The whole thing costs about four dollars a day in compute.
Worst case, I lose some money and publish an interesting post-mortem. Best case, the bots find an edge and I publish that too.
Either way, you can watch.