Built an AI trading bot that uses Gemini vision to look at actual charts before taking trades - not just raw price data. The core idea: generate a 1H+4H chart image, feed it to a vision model with context (funding rates, S/R levels, calibration notes from recent closed trades), get a score 1-10. Score >= 7 fires the trade. What surprised me: Gemini actually gives nuanced rejections - "4H downtrend, this looks like a dead cat bounce, score 4" - not just rubber-stamping entries. Early results: 33% win rate over 3 closed trades (small sample). The AI pre-filter cuts ~70% of signals to keep API costs in check. Full writeup on Stacker News: https://stacker.news/items/1449381 Has anyone tried vision models for chart analysis? Curious what approaches have worked. #bitcoin #lightning #AI #trading