A live US-stock scalping analyzer in Python, run against a local Moomoo OpenD instance with yfinance as the fallback feed. It pulls 1m/5m/15m candles, computes EMA/MACD/RSI and session-anchored VWAP with pandas + ta, and checks seven classic setups — opening-range breakout, VWAP bounce, bull flag — with a confluence checklist and a 0–10 safety score that penalizes lunch chop and fading momentum. Wired up as a Claude Code skill, so "check NVDA" returns an entry/target/stop card. The pipeline is deliberately linear — feed → candles → checks → score → card — so when a reading looks wrong there is exactly one stage to inspect. Analysis only — every order is placed by hand.

• Python
• pandas + ta
• Moomoo OpenD
• yfinance fallback
• Claude Code Skill

My biggest personal build: an AI trading assistant living in Telegram — python-telegram-bot at the front, the OpenAI API behind a multi-agent pipeline — analyzer, technical analyst, strategy generator, backtester, final decision — with Finnhub and yfinance for market data, a Streamlit backoffice, deployed on Zeabur. The agents run as a fixed chain — each one's output is the next one's input — so a verdict is traceable back through the backtest, strategy and analysis stages that produced it. Strategies live in a JSON skill library: the orchestrator matches them to market conditions, ranks them by tracked performance, and feeds trade outcomes back in. Every answer is forced into a tight format: verdict, entry, target, stop, one reason. About 8k lines of Python.

• python-telegram-bot
• OpenAI API
• Finnhub
• Streamlit · Zeabur

No longer a blueprint — a Minisforum UM890 Pro (Ryzen 9 8945HS, 32GB) now runs 24/7 in my living room: Windows 11 with WSL2 Ubuntu, Hermes installed bare-metal inside, plus a FastAPI trading server and a static daily-news page. Access is Tailscale-only — zero open ports, with a Windows portproxy bridging into WSL. The honest lesson: local models were the plan, and they lost — Ollama running Qwen 2.5 7B and Llama 3.1 8B couldn't keep up with multi-round tool calling on CPU-only inference, so a tiered router now sends routine work to budget cloud models and reserves a heavyweight for deep reasoning. The security model survived contact intact: sandboxed agent, outbound allowlist, low-privilege accounts per service, and a green/yellow/red delegation table — banking credentials stay red, forever.

• UM890 · WSL2
• Tailscale-only
• Ollama → tiered cloud
• 24/7 since May 2026