Our brain can estimate something and learn to be better at it next time. Intuition I think is the best weapon against estimating something. The surprising fact is that we wouldn't be completely aware of all the variables it considers while coming up with estimation. In the end it's all guessing. Planning is guessing, do it more often! The actual question one should ask is how frequent should you estimate. At dropx.io we follow a hybrid model of kanban with scrum meetings. It took me approximately 2-3 weeks to get an intuitive way of judgement for estimating a particular feature. These days it's easy to guess, but a deductive way of estimating would be to calculate individual velocities of all the team members and then we can definitely end up with a calculated guess(estimate).

If the question is taken literally, a perfect solution would be designing a swagger spec in YAML using swagger(now OpenAPI) standards and using any third-party service to create and host the application. RAML would also be another option here. In Python there are several libraries, some of which even can open a simple single call API in 3-4 lines of code -- quick enough? If we'd have to choose between REST API frameworks to develop an application with heavy business logic, in my opinion Loopback is the way to go. It will help you keep the API completely in compliant with the swagger specs. It can automatically generate the YAML file if you require. All you have to do is create data schema in JSON. It can generate client SDKs, for example the a services.js file would be generated if in case of angular. Lot of popular MVC frameworks like Django added additional libraries for generating REST APIs. But the middleware involved in the request-response cycles is still there. That'll definitely do unnecessary work at times and increase latency. I prefer to stay away from them than digging deep to disable the middleware.

A few things to look at when implementing any ML related stuff: Availability of Basic Libraries : It's very important to have a few core libraries like scipy/numpy/scikit-learn available when starting. The vector/matrix based calculations will consume lot of time, if you start from scratch because there are hardly any libraries available in JS. But this applies only if we use ML for some application oriented use case. Ignore this if you want to just work in machine learning for research purpose. Peculiar features of JS : When working in JS, we can think of putting features like async etc. to work. I'm not quite sure about any specific ML algorithms that would make use of these. But you can look into those. Thinking in JS : I found it a little hard to think in JS when implementing a few algorithms like tf-idf. Basically the non-blocking way of coding was what I felt a bit different from what we do in Python. Community support : There are hardly any programmers implementing ML in JS. You might not get enough support, which is essential when you do it from scratch. But there are a bunch of small libraries, unfinished projects which basically were trials, but I see no activity there. And there are lot of assumptions those developers made when compared with the original algorithms. So, I wouldn't suggest you to start working on top of them. Anyways, I'm not completely sure what you do use it for. These are my experiences. Hope it helps you decide!