Designing spectrometer is what I do for a living and with my experience and knowledge, I have serious doubts this device has sufficient resolving power to do what they claim it can/would/should do. To identify chemical components, you need a minimum spectral resolution (depending on the species you want to identify). To do quantitative analysis, the requirements are event higher. Typically, for solid NIR spectroscopy, I would aim at 2 to 4 cm^-1 spectral resolution. Under this, you can maybe check for the presence of a specific compound or compound family, but the capability to do so will be very dependent on the overall chemical composition.
Its possible to reduce the size of a spectrometer while somewhat keeping the resolution. But that goes only up to a certain extent... and that goes only with trade on signal to noise ratio. At some point physics overtakes wishful thinking. Reducing the instrument, and thus the optical throughput, you need longer measurement times to achieve adequate signal quality. Quantitative analysis with a (large) lab NIR spectrometer can take minutes, depending on the material being analysed. When you design spectrometers, you are constantly trading on aspect for another and by bringing a NIR spectrometer to that size, you traded a LOT of stuff away.
I also see spectral calibration being an issue with this device, then it works in reflectance and not in transmittance. It cannot be self-calibrating and directly provide a transmission/absorption spectrum. Maybe it is calibrated once during the production and assumed to be stable? If that is sufficient is, from my experience, questionable.
On the other hand, this is a very exiting breakthrough. I might even get my hands on one for fun. Why? because its, as they so well market it, a liberalisation of matter. Its a first step in being able to identify any substance that we get our hands on. While it may not yet be able to provide a full chemical make-up of a product, with enough a priori information it may be very useful.
Let me give you an example where such a device can be its money worth. When you buy fruits and vegetables that are bio/organic, you want them really to be that way. This decision to spend more money on these healthier food items is solely based on trust, which is often exploited. I doubt that the analysis of such a product can do what they claim (most of the return information is most likely deduced from the a priori information provided). But even with a limited spectral resolution and sensitivity, it should be able to identify spectral signatures of typical herbicides and pesticides.