I'm about to sleep, so I'll have to be brief, but I thought I should say at least something back, since you make some good points.
I'm not trusting their spectra on the kickstarter page. Frankly, I know they're claiming that's how it works, but I've no way of verifying or disproving that in any way, it's not really a testable claim right now. They're saying "Yeah, this is a spectra", but we've no way to distinguish that from, say, a bunch of lines drawn on a square in photoshop, especially with no labels on that group of coloured lines, or a spectrum taken from other equipment.
They're definitely not using Raman Spectroscopy. That would require either a laser(and a decently powerful one to boot), or a light much, much brighter than the device seems to be equipped with - as in, bright enough to be hazardous to the eyes. Though, I might be wrong on that point - I will probably end up saying this many times over this, so pardon me for going all broken record on you - but I'm not a scientist, I'm a journalist, and doing my homework is no replacement for years of education and experience.
They don't make the explicit claim they can identify individual molecular species out of millions(or any other number, for that matter), but the claim is strongly implied, and I would even say necessary for the device to operate. For example, their Tylenol pill - to take an example from the kickstarter page of something it can identify - they have to be able to distinguish corn starch, magnesium stearate, powdered cellulose, sodium starch glycolate and Acetaminophen from each other, and make a decent guess to their amounts to identify the product. Even in a well-equipped lab, that isn't just a done-in-one task, to positively identify all that, you'd need multiple types of spectroscopy(If you didn't want to use some other method, but let's say we don't, considering) and experienced, educated people to be able to filter and interpret the data correctly.
I mean, look at the techcrunch disrupt video. They supposedly differentiate between two different types of cheese, as well as their fat content and caloric content from their specific fingerprint. Now, I'm admittedly not a biochemist, analytical chemist or any other such thing by trade, but every one of them that I consult(Point of order - I've been on the trail of this for about a month now, after the campaign started popping up around the internet, so I've had a little time to work on this, as much as I wish I was magic enough to pull together a crew of scientists at 11:44 pm to look at kickstarter campaigns) notes that this would be an incredibly difficult thing to do with only a single NIRS analysis, and frankly wouldn't currently be possible with the technology we have.
That doesn't rule it out, of course, but it does make it very unlikely - returning again to the point that they would have to have made these incredible leaps in technology without making even a ripple in the metaphorical pond of instrumentation research and development, or any other related field.