IT professionals acquire AI expertise in a short time by attending couple of online courses. Once the courses are completed successfully, these new gen AI experts look for opportunities to hit problems with the newly acquired AI hammer. It is a challenge to identify problems that need AI for a solution. Majority of the problems seen as an AI problem can be solved easily by traditional statistical methods. But the IT child with new AI hammer in hand thinks that all problems in this world are AI problems and can be solved only by applying AI techniques.
Fraud detection, Customer segmentation, Churn analysis, Risk Analysis, Recommender systems, Binary decision making etc are commonly solved problems using AI and multiple solutions based on a variety of machine learning algorithms are openly available for these problems. If you are lucky to identify a customer looking for a solution in these domains, you can pitch in and design a solution by leveraging from the existing solutions. In AI world, solutions exhibit similarities and they can be generalized to create products. There exist challenges in bringing out products that can replace a set of AI solutions having similarity. The major reason for this is that the AI solutions are data dependent and when a generalized product is to be developed, the system has to be trained on data sets from all domains in which they are getting used.
Data is the fuel for AI solutions and getting access to large volume of training data sets from multiple sources is always a challenge. Unless the AI system is trained with multiple data sets, the AI product fails to give a meaningful solution to problems. In general, the attributes of good AI solutions are:
– Start with a strong business problem that cannot be solved using conventional IT using statistics.
– Provide excellent customer delight when they look at the results
– Provide measurable business value that can appreciated by the users.
For example: Percentage increase in the volume of sales.
– It should be fairly easy to explain the results produced by the AI system to the end users.
– The outputs generated by AI systems should be free of biases.
– The system should be capable of performing continuous learning from the newly arriving data. With the acquisition of new knowledge, the system should give consistent results with the results already generated in the past.
To begin an AI project, you need to translate a business problem into a machine learning problem and collect large volumes of past data with many attributes that can be used for training the models. Sometimes, you may not be able to get hold of training data existing in directly usable form. In this case, you may have to engage someone who provides data labeling services. The quality of the deliverable of AI solutions depends on the quality of data labeling. You should always look for someone with good domain expertise for data labeling. An expert service provider in data labeling in one domain may fail miserably in labeling data in other domains.
And finally, if someone is looking for AI solutions and did not have any past data, you should educate him with the need for past data in training the machine learning models. May be you can institutionalize a process for preserving relevant data and start the AI project when sufficient data becomes available.
If you are planning for a cloud deployed AI product, you will be able to subscribe and use pre-trained models exposed in the form of APIs by the cloud service providers. For example, Google Cloud Platform (GCP), Amazon Web Services (AWS) and Microsoft Azure provides several APIs for common problems like Vision, Speech, Forecasting, Classification, Recommender systems, intelligent user interface, natural language understanding and machine translation between languages. Most of these APIs are trained on huge volume of data and gives results with highest accuracy in real time. Also, you have the choice to make use of pre-built jupiter notebooks provided by cloud service providers for common problems. For AI solutions addressing specific issues of corporate clients, pre-built APIs are not yet available. Going forward, an AI product will be making use of home grown models along with standard APIs provided by cloud service providers. So, as an AI product developer, you need to learn and closely understand the capabilities of readily available APIs available with the cloud service providers. In near future, you will be able to develop AI products and solutions by stitching together a set of APIs provided by the cloud service providers with your own model. Understanding these APIs and the pre-built Jupiter notebooks will help you to save learning time and focuss on developing AI systems specific to the problem which you are addressing.
Feel free to learn the AI APIs available in the cloud. Wish you success in visualizing and developing a winning AI product.
See you next time…….
Janardhanan PS
Machine Learning Evangelist
