Building a product from scratch involves a sequence of processes starting from preparing a framework and applying the findings to design a model to iterating for the final result and putting it in a data pipeline. 

At MLDS 2021, Bragadeesh S, Head Of Analytics at Daimler AG, gave the audience a walkthrough on building a product using machine learning and AI, key skills required, challenges to overcome and more. Bragadeesh specialises in leveraging analytics to build products. The first step, according to him, is to lay out a concept. The next steps include building on it using AI and analytics and showing the product to consumers for their feedback to improvise it further. 

Methodology

He divided the entire process of building a product using ML into the following steps:

• Define the problem and arrive at MVP
• Gather data and prepare the methodology
• Choose the framework and right models
• Integrate and make it sustainable
• Deployment and monitoring

Define a problem and arrive at MVP

This step involves choosing the right objective for developing a product and identifying the problems in creating it. Once the first draft of the product or the Minimal Viable Product (MVP) is developed, the next step is to run it to check for problems. The next step is to choose the metrics wisely and appropriate approaches to solve the problem. 

Data gathering and preparation

Bragadeesh said choosing the data that can prove or disprove the hypothesis and problem statement is essential. For that, you have to identify the data source and create a repository first. Explore the data and choose the right column and fields. He advises fetching data either through APIs, big data methodology or from DWHs. The next steps include cleaning and processing the data followed by storing the cleaned data.

Model selection and finalisation

The next step is model selection and finalisation, where relevant frameworks are developed to help build solutions. Here models will interpret the data and output the result. Exploratory analytics — both univariate and bivariate — should be carried out to summarise the model’s main characteristics. Next steps include feature engineering and feature selection to extract features from the data. Finally, model selection can be made using ensembling methods, for instance, neural networks. Lastly, model finalisation can be done by serialising the data for which tools such as HDF5 can be used. 

Integration

This step involves bringing all the pieces of a puzzle together to solve the problem. In a nutshell, it involves steps such as choosing the framework (using flask, Django, streamlit, node J5), building a user interface (using HTML, CSS and Bootstrap) and integrating it with the framework chosen — front-end, back-end, APIs and DBs. The final steps entail testing the components and validating the results. 

Deployment and monitoring

The final step is deploying the model. The solution developed should be sustainable and self-sufficient. Bragadeesh proposed contenarising through dockers for easier deployment. It can be deployed as a web app on a cloud or on-premise. Placing a monitoring mechanism is an important step in the product development strategy. For monitoring, implementing security measures is essential. Think about fall-overs in the event of disasters, advised Bragadeesh.