I built a compact pipeline to calibrate option models—Heston, Bates, and Kou—directly from implied-volatility surfaces. A small CNN maps total-variance grids to parameters; I then reprice the full surface using a fast, stable Fourier–COS method. Two practical guardrails make it work: pricing on forward moneyness and clamping option prices to no-arb bounds before IV inversion. Training uses z-scored targets with constrained decoding. Diagnostics compare true, label-repriced, and net-repriced smiles with per-tenor RMSE and ATM bias. Simple yfinance backtests show model strengths vary by regime.