Happens. Things like that happen when somebody works.

And we should also keep in our plan the delay of major effort required simply because we miss the right road from the beginning, and discover it when it’s too late, so rework again.

This happens because we’re sometimes greedy, and we want to minimize our price on volume without considering the effective price of human work, sometimes gabby, and we under evaluate the difficulty of certain technologies, the market requires us to use.

The development of a board starting from 0, requires a team of skilled persons and several quarters to be prototyped, debugged, certified and so on. Plus the charge of an industrialization and supplier process able to realize thousands without problems. More than one year’s of pain, trouble, and difficulty.

Also if you start from a SOM, with the idea that the most complex job is done, you are buying into a false premise. The routing of a DDR3-4 memory may be a complex thing, but not less than routing and driving a USB3 or Ethernet. From a HW point of view it is the same impedance checked net, but you gain a lot of dirty jobs to configure the GPIO as you like. And then go on with kernel configuration, driver installation, debugging, and certification and so on. As in the last paragraph.

The only real way to go, in a few months, to a working proto and another few months to a safe industrial production run, is to use an SBC already done, flexible enough to solve your specific problems, where all has been done and nothing more is to be done. Or just a very small part, hopefully by somebody else.