Well the first optimization would be to roll my own libc (not that hard at all :) ) where i would load api calls by ordinal and only those that i really need (ok, for the sake of compatibility, i would dig them from IAT table based on 16bit hash (folding fnv32) of API name + dll name). Same for all other libraries, everything compiled with agressive optimization (at least /O2). Next step would be to take compiler that is minimizing the bloat, tinyc (
https://bellard.org/tcc/), compiling 32 bit binaries just to save some space. Maybe even go for .com to avoid PE header bloat. At the end compress everything with something like upx, but probably i would roll my own PE compressor. Instead of using functions, the macros would be used, absolutely no classes, #pragma pack(1) all structures (i never tryed what tinyc does :D). Also merging PE sections will save some bytes.
Size optimizing is fun and you can learn a lot but it is dying art, probably 99.9999% of todays developers dont understand what I have written in first part (today, you are learning the programming, but very indequately what the OS does, actually typical today programer understands the programing but is clueless what his code does on low level) ... but +1 for anyone that goes into that direction, my boss at my first job was saying that good software fits to one 1.44 floppy but this is today violated by HHLL. Well business wise no need for that unless you are making malware, but still cool.