Owl

The problem

Scientific computing needs dependable building blocks: efficient numerical operators, a way to express and differentiate computations, and a path from model definition to execution. Owl brings these capabilities to the OCaml ecosystem.

My contribution

• Framework design: co-designed a computation-graph abstraction that separates definition from execution, supporting lazy evaluation, optimisation and compilation.
• Core implementation: worked on automatic differentiation and linear algebra, and implemented C convolution operators with cache-aware optimisation and parameter tuning.
• Memory and interoperability: led graph-level memory reuse, ONNX export and validation with ONNX Runtime; led the Zoo deployment framework with JavaScript and MirageOS backends.
• Maintenance: project lead and release manager since 2020, working on code review, issue discussions, CI/CD, releases, contributor coordination and documentation.

A measured result

Graph-level memory reuse reduced Mask R-CNN CPU inference memory by 86.6%, from approximately 6.5 GB to 0.87 GB, compared with the same framework without graph optimisation. Runtime, including graph construction, was reduced by 24.2% in that experiment. These are CPU inference results for the reported configuration.

The work is described in Parallel and Memory-Efficient Distributed Edge Learning in B5G IoT Networks.

Making the framework usable

I led the official tutorial, developed examples across numerical and machine-learning methods, and co-authored two books about Owl. The project received two grants from the OCaml Software Foundation supporting maintenance and open-access publication.