I think this approach inevitably results in too many tools, too few employees being expert in a significant percentage of the stack, high turnover from a fragile stack, and even higher than necessary engineering costs. Why not something simple like tRPC?