I have no idea if any of these would be successful, but they are interesting to me (not interesting enough to pursue though) and maybe someone wants to run with one of them.
Deep Neural Net based weather prediction
Maybe someone is already working on this, it seems obvious to me. My understanding of how weather prediction works is you have a grid on which you discretize the actual state of the world, then you have a physical model that you simulate to see how that state evolves over time. Do this a bunch of times with different random values perturbing the system and you can estimate the probability of the weather over the next x days wherever in the world you want. The quality of these predictions is obviously going to vary by how good your data is and how good your model is. Physically modeling the weather has a lot appeal from a scientific perspective, but in other domains statistical and ML based models often predict better than physical ones. Since DNNs are good at finding structure from data, and since we probably have pretty good daily or maybe even hourly weather data going back decades for most of the US, it should in principle be possible to train a DNN on all that data to predict future weather. If you did this how would it compare to the best physical models? One downside of these nonphysical models is they can predict obviously nonsensical things, for example, maybe your model says it will be sunny in a region that is completely surrounded by a storm system. But these sorts of problem can probably be fixed by rather simple rules that cleanup the output.
Better dating app based on personalized matchmaking
If you are an unattractive man dating apps aren't a great way to meet women. If you're a plain looking woman dating apps aren't a great way to meet a man for a long term relationship. The problem with apps is they place too much value on looks and not enough on personality. What if instead of seeing what someone looked like you just got matched with them, got a brief bio about them, and had to decide whether or not to go on a date. Probably you'd need some sort of physical matchmaker to interview people (probably over video call), rate them on attractiveness, assess their personality, and somehow pair them with someone else. You might object that this is too subjective, but people deciding for themselves who to go on dates with is pretty subjective too and most of them aren't very good at it. It seems likely that a skilled matchmaker could do better than many (maybe most) people do for themselves.
I see a couple of potential problems with this sort of app. First, the highest value type of man probably wouldn't switch to it. Apps work fine for them. Second, it's possible women wouldn't switch either. They may prefer the reality of a fling with the top echelon of men and the (possibly imagined) possibility of that turning into something more, than the more practical goal of a long-term relationship with someone they are more compatible with long-term. Of the three ideas I've proposed here this is the one I'm most skeptical of. Revealed preferences show women and high value men are pretty happy with the existing set of apps. I could imagine a total lack of product market fit even if this idea were perfectly executed, but it still seems worth trying.
Political campaign managements as a service
Dominic Cummings and Vote Leave had a sophisticated system for crafting their political messaging and determining what voters to target in order to swing the vote in their favor. Regardless of what you think of Vote Leave's political goals, they were effective. Someone who could build a similar system in the US would be a formidable force in the US political environment. If this system worked well, the mere threat of using it in support of certain candidates would be enough to force otherwise impractical ideas through the gridlocked congressional system. Of course, any successful system would be imitated by others and you'd face something of a Red Queen problem, however, the failure of Republicans to respond to Democrats successes with mail-in voting shows that simple political advantages can persist for longer than you would expect. This seems like a potentially very high-leverage problem that no one is seriously working on.