ai-aas

AI as a Service. Production is all you need

Acknowledgement

Google supported this work by providing Google Cloud credit. Thank you Google for supporting the open source! 🎉

Overview

In this project, I aim at developing an embarrassingly simple, production ready AI as a Service.

Motivation

There are lots of free and open AI/ML projects published by great folks across the globe. However, it is not that easy to integrate such models into existing projects and deploy them in production. Therefore, we need a simple way to run them as an API service on various deployment targets.

Design Goals

Ideally, this project should be:

• technically and legally ready for use in production.
• containerized and self-contained.
• configurable with simple config files and/or environment variables.
• cross-platform and accelerator-agnostic.
• deployable with only a few clicks and/or commands.
• scalable for extreme use cases.
• modular so that it can support choosing different sets of models.
• predictable and self-documented.
• maintainable with minimum dependencies.
• easily expandable with new models.
• usable for batch and/or online prediction.

How to use

This project makes use of Docker Compose Profiles to support optional enablement of services, so you need to have Docker engine V20.10.5 (or above) installed.

git clone https://github.com/monatis/ai-aas.git
cd ai-aas
export COMPOSE_PROFILES=zsl,ner
docker-compose up -d

Current supported tasks

• Zero-shot text classification
• Named entity recognition
• Question answering
• Question paraphrasing
• Question generation (?)
• Abstractive summarization

Very soon

• Text clustering
• Scalable semantic search
• And more

Plan of Attack

• Identify a few suitable models for the initial release.
• Identify the minimum dependencies (TensorFlow, FastAPI, Redis, Traefik and possibly a few more).
• Design the infrastructure.
• Implement abstract request and response layers.
• Implement dependable request and response schemas for different data modalities.
• Implement unified and reusable preprocessing components.
• Implement prediction workers.
• Use other great services such as TF Hub as much as possible.
• Make it configurable thanks to docker-compose profiles.
• Make it optimized.
• Power great projects that need production-grade AI/ML services!