To make the most of the state-of-art AI/ML model advancements in the recent months, Machine Learning Engineers must empower app developers to use these models as easily as possible. There are several gaps to making this happen at the moment.
Most modern software development has moved to the cloud native stack containing kubernetes, containerised applications and microservices. Machine Learning models continue to lie in Jupyter Notebooks in research mode and teams face significant hurdles in moving models to the production/scaling stage. At Nautical, our vision is to ease the deployment and integration of machine learning models in modern software stacks.
To do this, we must first provide easy deployment options for models on kubernetes and we plan to provide support for this through Helm and Tanka. Once users are able to deploy these models on their kubernetes clusters, these models can be trated like external APIs that the application will interact with. The next step is to provide SDKs for application developers to interact with these ML models with ease. Here lies another problem - some models take images as input, some take time series and others take audio input. Serializing & deserializing is an unnecessary pain that developers face while writing home grown SDKs to ineract with models. Finally, ML models can be slow, lack any telemetry out of the box and require special configuration options for better performance. To solve for these use cases, Nautical will provide SDKs in popular languages like Python and Java so app developers can focus on the business parts of the app rather than worry about model integration.