Liberal use of libraries written in c (e.g. pandas, pytorch, numpy), some use of cython (not in the current version, but I have done so), and relying on time frames and strategies which have some tolerance for latency. If you trade five minutes after the start of a 1 day candle on the basis of where you expect it to be at close, it’s not such a big problem.

It’s a losing game to try and out-pace the big end of town.