My gold-trading automation is already 50 % finished, but the critical “last mile” still misbehaves. All core pieces are in place—YouTube monitor, Telethon listener, and an OANDA back-tester—yet real-world runs expose two stubborn faults that appear randomly during normal usage: • the YouTube module sometimes misses or mis-parses Entry / SL / TP whenever the channel goes live • the Telethon listener posts the same Paradise-channel trade more than once Here’s exactly what I need from you: 1. Harden the YouTube Live watcher – Detect every new live session from one specific channel via Google API v3, rotating keys as limits approach – Stream descriptions and live-chat in real time, then run the text through Gemini (or another LLM if you prefer) to extract the three fields with rock-solid accuracy – Return a clean Python object so the rest of the pipeline can consume it immediately 2. Repair and de-duplicate the Telegram listener – Work inside my existing Telethon script that already monitors five channels (three of them Paradise) – Ensure each signal is published to my downstream queue once and only once, regardless of edits or reposts 3. Validate each incoming signal against OANDA – Grab the last six months of 5-minute XAUUSD candles on demand – Compute win-rate statistics for the proposed Entry / SL / TP before the trade goes live – Fail gracefully on “peer/entity” disconnects and reconnect automatically Acceptance criteria • Zero duplicate trades from any monitored Telegram source over a 24-hour soak test • 100 % capture of Entry / SL / TP from at least ten consecutive YouTube live streams run back-to-back • No unhandled exceptions after a continuous 48-hour run with forced API-limit throttling The stack is pure Python 3.11 with Telethon, google-api-python-client, openai-python (for Gemini), pandas and a lightweight FastAPI endpoint for status checks. I’ll share the private repo and current .env once we kick off. If bulletproofing scrapers, juggling API quotas, and blending LLMs with market data is your idea of fun, let’s get this bot over the finish line.