AI and machine learning models are being used to help companies stay competitive by discovering new revenue opportunities, improving risk management, detecting fraud, and streamlining business processes. But years ago, data science wasn’t even on the curriculum at universities, so many software engineers are acquiring the necessary skills on their own.  

From my experience, anyone who has a strong STEM background can have a smooth transition to becoming a data scientist. Personally, I studied biology at university, but I was comfortable learning how to build machine learning models on my own. So, even if data science wasn’t what you studied in undergraduate or graduate studies, it’s possible to make the transition to AI and bring the power of machine learning to your teams. 

Here are some tips from my experience as a data scientist:

1. Study online: Those who want to start learning AI have plenty of online options available. Some of them are aimed towards people who already have a certain level of technical knowledge and focus on coding, while other courses are for those who don’t have any prior expertise in programming and engineering. If you’re an absolute beginner, start with some introductory Python courses and when you’re a bit more confident, move into data science, machine learning, and AI. Math is an absolute must for data science so it’s also a good idea to brush up on linear algebra, calculus, and probability and statistics. Entering a Kaggle competition can also be a great way to fine-tune your skills.

2. Assess the data situation: Companies commonly use analytics and machine learning for one purpose – to help improve the bottom line through increasing revenue and reducing costs. But then to quote the adage “garbage in, garbage out,” you can’t improve operations if you don’t have regular measurements over that time that can give you enough information to glean meaningful insights. One of my favorite data science jobs was with a site used to book hotel reservations, but I discovered that the data that was collected was limited to individual transactions without monitoring web navigation. We didn’t begin an AI project until we had gathered data for an entire year so that we had a strong basis to perform data analysis.

3. Choose a simple goal: It’s very important to choose a simple goal that can be communicated in a single sentence. After we collected data for a year at the hotel booking site, we chose to discover which hotels were chosen more often and then make them appear higher on the search list to increase the likelihood that a hotel will be booked on the company’s site. The project was a success because the goal could be easily explained, executed, and measured. Most importantly the results had an immediate positive impact on the bottom line. This is the most important tip for aspiring data scientists: Make sure that the enterprise can see the value of AI clearly and as soon as possible. 

4. Choose a platform based on the data: Today there are many different types of data including time series, tabular, natural language in many forms, images, audio, and video. It’s important to assess the long-term requirements over the lifetime of the problem addressed by the model to make sure you have the right technology to ingest and process all the different possible inputs. For example, medical machine learning projects can be the most complex because they involve medical text, scanned documents, hand-written notes, graphs, and medical images.

5. Build an end-to-end solution: Building a model that can generate insights is not enough. The model needs to be monitored for accuracy, and the pipelines need to be maintained to ensure a steady stream of quality data. One of my first machine learning models was managed by a home-grown MLOps system that had some hiccups. For a whole night, the system was generated bad hotel recommendations, including two-star hotels that charged over $300 a night, but we couldn’t identify and fix the problem until the following morning. Every machine learning project needs to address the long-term requirements by having a framework to retrain and test and roll back and start over if there are problems. 

6. Open source doesn’t have all the answers: It is always a good idea to see if someone has already developed, trained, and tested a model that can be used so you don’t have to reinvent the wheel. However, from my experience, many marketing, finance, and manufacturing departments need to build models trained on their own data. For example, one of my projects for construction safety needed to identify whether factory workers are following safety regulations by wearing helmets and gloves. There were open-source object detection models available, but they weren’t useful because they could only detect a limited array of objects in a narrow set of environments. 

7. Prepare to compromise: There is no perfect machine learning model, and the decision as to which model to use needs to be based on the financial impact. For example, I evaluated two fraud detection models where one was more effective at detecting large-scale fraud and another could identify smaller financial losses. A certain set of flagged transactions would also require some human review from a team of fraud analysts. We did a simple calculation to pick the model that would have the most positive financial impact and could produce a manageable number of transactions to review.  

8. You may become the victim of your own success: At one company the success of our project created a long list of internal departments who also wanted machine learning added to their IT agenda. Helping enterprises transition to becoming AI-first means educating developers and users to the importance of data, and then working with them to select problem that can be solved by AI in simple and explainable ways. It can also become a great opportunity to cross-train additional teams to start adopting ML into their workflows.

Each AI journey, for individuals and enterprises, is highly personal, but certain guidelines can ensure a smoother ride. AI has transitioned from being a technology of the future to a must-have to stay competitive, and I predict in the very near future that generating machine learning models will become a required skill for many developers and software engineers.