Kelvin Ma is a staff software engineer and product-focused technical lead whose work at Google Photos helped bring Magic Eraser and Magic Editor to an application serving more than 1.5 billion monthly users. A founding engineer on the team behind Photos’ machine-learning-powered editing features, he builds creative tools that make advanced image manipulation accessible without sacrificing reliability.
Originally from Canada, Ma studied computer science at the University of Waterloo, earning a Bachelor of Mathematics, and worked at Codecademy before joining Google. His professional website also describes graduate study in machine learning and computer vision through Georgia Tech.
At Google Photos, Ma helped develop on-device computational photography capabilities including portrait segmentation, adjustable background blur, lighting corrections, and image enhancement. The underlying architecture combines a shared C++ foundation across Android, iOS, and the web with mobile inference through TensorFlow Lite, now LiteRT. Shipping these features required controlling application size and model downloads, accommodating differences between devices, and addressing errors such as segmentation masks that miss fine strands of hair.
Magic Eraser, introduced in 2021, combined distractor detection, segmentation, image inpainting, and custom graphics rendering. Magic Editor subsequently expanded into cloud-backed generative editing, enabling people to reposition subjects, transform backgrounds, and remove objects alongside their reflections. That architectural shift introduced accelerator capacity planning, network latency, larger-model testing, and trust-and-safety constraints absent from device-local systems. Google later incorporated AI suggestions and targeted edits into a redesigned Photos editor.
Distinctive engineering principles
- Design around concrete user intent. Ma starts with recognizable tasks—moving a subject, removing a distraction, improving a background—and favors tapping or selecting an image over expecting people to write elaborate prompts.
- Treat evaluation as production infrastructure. His approach to representative model evaluation treats realistic benchmarks as the machine-learning equivalent of regression tests: they must capture actual user behavior, reveal unpredictable failures, and run quickly enough to improve products before launch.
- Choose architectures for practical reliability. Local inference avoids network delays and server costs; hosted models enable more ambitious edits but require capacity planning and different testing strategies. Ma favors smaller models, model distillation, or deterministic software when they deliver the same outcome more efficiently.
- Make creative uncertainty useful. When several plausible edits could satisfy the same request, offering multiple results can turn generative variability into a productive choice instead of pretending one answer is universally correct.
Beyond Photos, Ma created Reflect, a local-only journaling application incorporating language-model-assisted analysis. His GitHub profile and AI Engineer conference talk connect that project with his broader interest in practical, user-centered machine-learning systems.
