[1] https://github.com/karpathy/convnetjs [2] https://github.com/harthur/brain
Ironically, looking dennybritz's comment, JS is particularly ill-suited for this, as, as far as I'm aware, there are no libraries for working on the GPU, which is tremendously useful for working with machine learning algorithms which usually are crunching lots of data.
However, I don't know much about karpathy's specific implementation and experiments, so there might be a reason he choose JS. I'm thinking its more likely to make visualizations easier rather than any computational reason though, so you might be stuck in a niche there.
As for Karpathy's implementation, you're right, he chose JS for that reason. His library is meant as a tool for experimenting and teaching, not to deploy these models in production.
Op: Send me an email if you're interested in this (email in profile). I actually started something similar just recently at http://www.wildml.com/.
Be warned though, there are quite a few research papers, especially those coming out of Academia as opposed to industry, that are cryptic, lacking in detail, or not easily reproducible. And that's intentional. This isn't obvious when you read a paper, it will only become obvious when you actually try to reproduce it. These papers are the result of misaligned goals Academia in, where it's all about publishing as many papers as possible that look good on the surface. When In doubt, I'd recommend focusing on research that has a lot of interest from the industry. That's usually a safer bet.