I think that toolchains should have better support for estimated vs used time and then learning from that. You could classify tasks by field (DB, frontend, etc) and then calculate correction values based on past failures. Learning from wrong estimates seems to be even harder than estimating, because it hurts to admit that you were wrong, so some automated, blame-less version of it (per sprint or so) might be a good start.