A year building a Transformer + Mixture-of-Experts model for crypto microstructure trading on Bybit. The architecture, the training pipeline, the diagnostic experiments that ruled out the strategy class, and what I came away with.

The transferable methodology from a year of ML trading work: feasibility math, normalization and clipping, walk-forward validation, Monte Carlo, training/serving parity, signal filtering, kill-switches, and systematic rejection.