Latest idea I already posted about a few days ago: a keycard based doorlock with actual security (ed25519 based authentication, instead of the standard crackable 48-bit DES). Nobody really cares about that level of security.
A phone app that makes scanning documents stupid easy by using machine learning to identify when a document is on screen and automatically take a picture, and then use more machine learning to convert the picture into a cleaned up image. The concept to train the machine models is pretty easy; just print out a known document, take a bunch of pictures, and train the model to convert from the pictures to the original document. But it would likely require massive kernels to get enough context to do the necessary 3D transformations. Could probably reduce the strain on the algorithm by using traditional picture->document algorithms and then use the model to clean it up further. Either way, it's new tech that would require a lot of work.
Vocal isolator for extracting vocals from music, using a machine learning model. These tools kind of already exist, but they don't use machine learning, usually require user interaction to tweak knobs, and don't work particularly well. My bet is that a 1D conv network can do great, and generating examples for training is super easy. But, again, new tech, and as easy as machine learning sounds it usually devolves into long nights of debugging, tweaking, and banging your head trying to figure out why changing one hyper-parameter ever so slightly completely changes your results.
A hardware password manager. Like KeePass/LastPass/etc but as an actual hardware device, possibly a wearable/watch. The idea is that a hardware password manager provides more security than a software password manager, by virtue of not being susceptible to malware (though yes passwords can still be stolen 1 by 1 if used on an infected machine), and most importantly works with phones/tablets over bluetooth. Requires developing hardware, productization, etc. Not a big market; nobody cares about security.
I have a quiz website that I built for my wife, who's a teacher. It allows here to give online fill-in-the-blank type quizzes to her classes, with a set amount of time to take the quiz, and a set amount of time that the quiz is available. She's happy with it; I made it because she had something similar with Edmodo but the school dropped Edmodo. I can't imagine such a simple website is worth much though, and it's currently only designed for her and a bit rough around the edges. It'd take quite a lot of UI updates and a bit of backend tweaking to make it usable by the general public.
A couple game ideas, mostly geared towards programmers. Pretty niche, and games take immense effort versus profit.
A service to provide email to Bitcoin addresses. You authenticate yourself using your Bitcoin wallet's signature mechanism, and then can receive email like at e.g. 1p7H5w1LfgLT6tat951e85dFXEDQjNp8L@example.com. I got this mostly working, but rendering emails is super hard, who wants to use yet another mail client, how do I monetize, how do I fight spam, etc, etc. Lots of security and anonymity angles to this, but security is hard, so again lots of work.
NSFW: Machine learning based porn assistant. You thumb up/down images, it learns what you like, and finds more of it. Keeps learning as you thumb up/down its results. Classic machine learning model. Requires new tech, and it probably requires training a model per user (super expensive), or some kind of transfer learning.
Lots more beyond that, all very pie-in-the-sky, new tech related, or super niche.