Enzymes are broadly used as biocatalysts in industry and medicine due to their coverage of vast areas of chemical space, their exquisite selectivity and efficiency as well as the mild reaction conditions at which they operate. Custom designed enzymes can produce tailor-made biocatalysts with potential applications extending beyond natural reactions. However, current design methods require testing of high numbers of designs and mostly produce de novo enzymes with low catalytic activities. As a result, they require costly experimental optimization and high-throughput screening to be industrially viable. Here we present rotamer inverted fragment finder–diffusion (Riff-Diff), a hybrid machine learning and atomistic modelling strategy for scaffolding catalytic arrays in de novo proteins. We highlight the general applicability of Riff-Diff by designing enzymes for two mechanistically distinct chemical transformations, the retro-aldol reaction and the Morita-Baylis-Hillman reaction. We show that in both cases it is possible to generate catalysts exhibiting activities rivalling those optimized by in-vitro evolution, along with exquisite stereoselectivity. High resolution structures of six of the designs revealed an angstrom level of active site design precision. The design strategy can, in principle, be applied to any catalytically competent amino acid constellation. These findings enable the practical applicability of de novo protein catalysts in synthesis and shed light on fundamental principles of protein design and enzyme catalysis.