01 · Work · Meridian Capital
A public experiment: an autonomous trading system where six AI agents research a stock, debate it, and place the trade. It runs on free LLMs, either Groq's free tier or a local Ollama model, so it costs nothing to run.
Everything happens in Alpaca paper trading. No real money moves. It defaults to paper and only trades live if you deliberately switch it to live mode and give it a live key. I built it to learn how multi-agent systems reason and disagree, and the whole thing is open on GitHub.
02 · Story
Why I built it
I wanted to learn how multi-agent systems reason, argue and veto each other, on something with a clear score: a portfolio against the S&P 500.
Who stays in controlPaper money only by default, and a kill switch halts all trading at once.
03 · How it works
One agent researches news and fundamentals, one runs the technicals, one reads the macro cycle, one manages risk, one makes the final call, and one places the order. Each stock flows through all six every cycle.
If the researcher's conviction is too low the pipeline short-circuits and skips the stock to save API calls. The risk manager can hard-veto a trade, and the portfolio manager cannot override it.
When a position sells, the actual profit or loss is written back as a plain-English lesson into a SQLite memory table. Next time the agents look at that stock, their prompt includes the last few lessons.
A Next.js dashboard streams every decision over Server-Sent Events: the agent debate, the portfolio, benchmarks against the S&P 500, and a manual kill switch that halts all trading instantly.
04 · Features
No single position can pass 15% of the portfolio, and it can never spend more cash than it has.
An 8% stop-loss is checked every cycle, and a 5% loss in one day trips the kill switch on its own.
The manual kill switch survives a code reload, so a restart cannot quietly switch trading back on.
The trader holds every position for at least 24 hours.
In a bear cycle the portfolio manager can hedge with inverse funds instead of only selling.
Each agent's accuracy is tracked on the dashboard, next to the S&P 500, NASDAQ, Nifty 50 and Sensex.
05 · Scenarios
Learning AI agents
Each decision shows every agent's conviction, the risk manager's veto, and the final call side by side.
Running it for free
Groq's free tier or a local Ollama model, on a free host.
Testing risk rules
The same limits apply in paper and live mode, so every rule is proven on paper first.
I am not looking for work. I am looking for good people to know.