Many engineers reinvent mathematical wheel when working with resource-constrained devices, unaware of proven efficient algorithms from computing's early days.
A library of well-optimized, clearly documented implementations of classic algorithms like CORDIC, with modern SIMD and GPU acceleration where applicable.
Embedded systems developers and low-latency application engineers would pay for production-ready implementations.
Start with the most requested algorithms from historical computing literature.
Risk is niche appeal - most developers just use standard math libraries.