Instead of dumping raw price data into a prompt, I treated the LLM strictly as a "Qualitative Reasoner" (the Brain) and built deterministic scripts as the "Quantitative Executors" (the Hands).
The Math Layer (Zero Hallucination): Traditional Node.js/Python scripts pull market data and calculate deterministic indicators (SMA, ATR, RSI) and backtest historical win rates.
The Context Layer: It fetches live news, SEC filings summaries, and Wall Street ratings via Finnhub/Tavily.
The LLM Brain (DeepSeek-R1 / GPT-4o): The model is ONLY fed the computed results. It acts as an investment committee, weighing the technical setup against the macro news catalysts.
The Risk Controller: Even if the AI screams "BUY", the system passes the signal through a hard-coded Volatility Parity sizing model (using ATR) to calculate the exact safe position size.
Human-in-the-loop (Telegram): For execution, I integrated a Policy Engine. The bot sends a beautiful proposal card to my Telegram. It only executes via the Broker API (Longport/Alpaca) if I reply /approve.
The $1000 Experiment:
I am running this live right now. It's not about getting rich; it's an engineering experiment to see if multi-agent systems can actually maintain a positive Expected Value (+EV) in noisy markets without blowing up.
Currently, the system is [up 2% / down 1% / flat] in its first week. It successfully avoided a fake breakout yesterday because the Bear-Agent successfully debated the Bull-Agent on low volume.
Tech Stack: Node.js, TypeScript, SQLite, better-sqlite3, Longport SDK.
I’m open-sourcing the entire framework. It’s highly modular, so you can swap in your own indicators or prompt logic easily.
I would love for the HN community to roast my architecture, point out my blind spots in the backtesting logic, or just tell me how quickly I'm going to lose my $1000.