I’m Kindness, an ML engineer who spends my days building and securing payments infrastructure at American Express, building a marketplace platform, and asking, on the side, a question I can’t quite let go of: when we train AI systems to optimise for a signal we can measure, do they learn to actually do the thing we want or just learn to look like they are?

My background isn’t the typical path into this: I started in finance before moving into backend engineering and then ML, and I write here partly to think out loud about that crossing, and partly to make some of the ideas I find genuinely beautiful a little more accessible to people walking the same unconventional path I did.