We report the development and application of a scalable machine learning (ML) framework (Minerva) for highly parallel multi-objective reaction optimisation with automated high-throughput experimentation (HTE). Minerva demonstrates robust performance with experimental data-derived benchmarks, efficiently handling large parallel batches, high-dimensional search spaces, reaction noise, and batch constraints present in real-world laboratories. Validating our approach experimentally, we apply Minerva in a 96-well HTE reaction optimisation campaign for a nickel-catalysed Suzuki reaction, tackling challenges in non-precious metal catalysis. Our approach effectively navigates the complex reaction landscape with unexpected chemical reactivity, outperforming traditional experimentalist-driven methods. Extending to industrial applications, we deploy Minerva in pharmaceutical process development, successfully optimising two active pharmaceutical ingredient (API) syntheses. For both a Ni-catalysed Suzuki coupling and a Pd-catalysed Buchwald-Hartwig reaction, our approach identifies multiple conditions achieving >95 area percent (AP) yield and selectivity, directly translating to improved process conditions at scale.