I applaud the approach, but it seems like a challenge to achieve useful accuracy across many devices. I am a little skeptical about the 3nm claims. First, the

I applaud the approach, but it seems like a challenge to achieve useful accuracy across many devices. I am a little skeptical about the 3nm claims. First, the (bandpass) Bayer color filter arrays on cameras limit their color resolution pretty sharply. It might be a better (if less elegant) approach to use the ambient light sensor and scan it across the spectrum generated by the diffraction grating. Also, the resolution of a spectrometer is in part determined by entry slit width. Old timers and DIY'ers use abutted razorblades to make a clean, straight slit. Problem here is that there is a trade-off. You need a wide slit to get light in for these insensitive webcam sensors. But then you lose spectral resolution because your slit is wide. Alternately, you narrow the slit to get more spectral resolution, but then your image is dominated by noise. I'm sure these guys have found the sweet spot. Second, the spectral response of the color filter arrays varies both from camera to camera and even within same-model sensors (I personally own over 40 of one model of Canon Powershot, and the differences in color response are fairly extreme). This could in principle be at least partly calibrated out, if you had each device take a picture of a calibration standard. Third, the image signal processors are closer and closer to the sensor these days. It may be impossible to get data which hasn't at least been somewhat mucked with, meaning some people will have a useful spectrometer and some won't. This goes doubly for phones with bad imaging pipelines, where the image processing is out of your control. Lots of spatial and spectral processing going on. It will be interesting to see, for example, how the auto white balance and sharpening algos deal with having zero natural image content in the captured data. Aside from the within-sensor variation/image processing problems it would definitely be usable for relative data - I see that they used a relative approach on their UV detection of a bluing dye in detergent. Nice. All that said, I run a somewhat-similar project using cheap cameras to do work previously only reserved for hard-core SLRs and scanning line sensors, and I've found that over and over again, something is vastly better than nothing and often "good enough" is better than expensive, complicated, and pain in the ass. I've also dealt with a lot of people laying out a bunch of complicated technical "Here's why it won't work" talk and have learned to counter with my own "But look, it does work" answer. Blah blah: I think it's important for the capture software captured as much metadata about the image sensor as possible. Their software looks great and a database of spectra would also be great. It would be AWESOME if the Shazam-style auto identification worked at all. Good luck, I'll probably be supporting the project.