Solo build — Tony (tonnylegacy18)

APEX turns Claude Opus 4.7 into an autonomous trading-strategy optimizer for MetaTrader 5. Most strategy optimizers are brute-force grid searches: pick a metric, sweep N parameters, pray. The user gets a winning configuration but no idea why it won, no confidence in whether it'll generalize, and no transparency into the search process. As a forex trader, I lived this pain running backtests, screenshotting results into AI chats to ask "what does this mean?", tweaking one parameter, and repeating for hours. APEX replaces the grid with an AI loop. Claude Opus 4.7 reads each backtest result, looks at the full iteration history, considers the parameter schema and quality targets, and returns structured JSON with concrete bounds-checked parameter values, per-change reasoning, a confidence score, and a goal-status breakdown. Every change, every reason, and every Claude token streams live to a dashboard so the user watches the AI think. The pipeline runs three phases: (1) Latin-Hypercube exploration to map profitable parameter regions, (2) autonomous AI iteration toward user-set quality targets (PF / drawdown / Calmar) with stuck-detection and random-escape, (3) out-of-sample + sensitivity validation → verdict (RECOMMENDED / RISKY / NOT_RELIABLE). The point isn't a better trading strategy, it's a working pattern for AI-as-driver of a long-running optimization loop, with reasoning fully visible. That pattern transfers to ML hyperparameter tuning, A/B variant generation, ad-creative optimization, and infrastructure cost tuning.