Real Time Wine Quality Model

A full, production-ready system which is designed to make predictions on the expected quality of the wine based upon fluctuating IoT readings.

Introduction

Welcome!

Thank you for taking the time to explore this repository! This project was developed to help new ML engineers wrap their head around some of the undertaught parts of computer science: developing a fully-functioning solution for your clients or employer. It covers less of the concepts around Data Science, such as model selection and feature engineering, and instead focuses on how you can deploy these models into a value-generating environment.

This system has been designed around an imaginary scenario. I've deliberately picked a trivial yet realistic example to help you understand why I've made the decisions that I did. The focus of this project is not

Business Context

You've been approached by a winemaker who is looking at employing machine learning in her business to help drive a more data-driven decision making process around his cask pricing. This winemaker has IoT sensors connected to her barrels that is constantly feeding information into her sensor database. She knows that this data being collected, fundamentally, captures the final quality of the wine that is produced. She isn't quite sure of the specifc values that makes a good wine and as such she thinks that machine learning might be able to determine what makes a good wine.

Currently, she has to taste every barrel of wine that's produced to grade it on a score of 1-10. This grade ultimately effects the price that she can sell the wine at. These tasting sessions take up 10 man-hours a week from her employees and are prone to subjective palettes rather than objective ratings. She'd like to, ultimatley, completely remove the need to conduct tastings to price her wines.

Technical Context

The IoT data is collected in a real-time fashion and is constantly updating the current reading values in her DynamoDB NoSQL database. You can expect that the readings will change on average every 10 seconds.

The interactions between features that make up a good wine are fairly constant, and as such an ad-hoc and manual approach to updating the trained model is sufficient.

Requirements

In order to get this solution working you will require the following installed into your system:

Terraform (>= 0.12)
Miniconda or Anaconda package manager
AWS Account with credentials

Installation

You can install the source code through the Makefile utilities. To setup the project run:

make init

This will install any necessary dependencies into your environment. You will need to build the source code and push it into your AWS environemt before you can construct the infrastructure. To do this, run the commands:

make build
make push

Once this is complete, you can standup a copy of the infrastructure using the command:

make infrastructure

This will create your own copy of the environment within your default AWS environment.

This infrastructure won't have any data pushed into it: you'll have to do this yourself. The command to do this is:

make database

Now you have everything to get your solution working! You can manually peturb the values in your database to see the reflected effects, or you can run another utility which simulates the real-time IoT data:

make simulation

Now you'll be changing the values of the wine data at a rate of approximately every 10 seconds.

Once you're done exploring the source code you can destroy all of the resources created using the command:

make destroy