Have anyone used machine learning in real world use cases? (would be nice if you can describe the case in a few words)
I'm reading about the topic and do some testing stuff but at the moment my feeling is that ml is like blockchain. It solves a specific type of problem and for some reason everyone wants to have this problem.

I have read an article of using ML to control a cat door. It was quite instructive and seemed very appropriate for that use.

Image or pattern recognition seems to be appropriate.

I haven't, personally (unless a StarCraft II bot counts as real-world), but I've seen several applications of machine learning in some pretty cool projects. It's being used in the medical field to detect and diagnose medical problems with unprecedented accuracy. They're also using it to generate new chemical compounds for medicine. They've also demonstrated the risks though, as one team of researchers flipped the success conditions on the chemical compound model and generated 40,000 chemical weapons in 6 hours.

Self-driving cars are being trained with machine learning (and reCaptcha data is being used to verify the results, I think, which is why they always ask you to identify traffic lights and cars and stuff).

Websites use it to power their recommendations. Phones use it for Face Detection so you can unlock your phone with your face. Environmentalists use it to model species populations. Machine translation services use it to improve accuracy. And I'm out of characters.

I worked on ATS(Application Tracking System) at my last company and we used ML to rate and recommend candidates.

Ember put up a good list, also there's the vague area of testing your product and just throwing that data in ML. I've seen examples of a) judging by just the amperage over time of a valve if a valve-action succeeded or not b) meassuring aging and having your ML then prognose or counteract that.

I find it downright amazing that before that you'd have to come up with a scientific / engineering model of what exactly happens why. Now you don't care and just skip that

My company has built a lot of ML models from clients from how to more effectively distribute stock to stores to drones flying around an oil rig or gas plant and taking photos of pipes and components to identify rust and using gas plant sensor data to detect a trip in advance so that engineers have time to correct the problem before the plant has to shut down for a few days.

I'm not involved, but our company has a long running project for predicting delays in public transport. The input is stuff like road work or holidays. The itineraries are "pessimistic", then the drivers use it to adjust/compensate by starting early or by waiting a couple minutes at each stop. The goal is to be consistent, not fast.

Yup.
I built a personal project to automate an old online Flash game. Used the project to learn a lot of things outside my comfort zone at the time. The baseline automation used SikuliX for 95% of the automation. But I had a problem with a random location popup messgae - I trained ML model to take a screen cap, and locate the popup for the s Sikuli script to click on.