“My only advice is “Avoid shortcuts”. You can’t leave statistics, maths and coding behind to move on with AI/ML.”Nitin Aggarwal
For this week’s ML practitioner’s series, Analytics India Magazine(AIM) got in touch with Nitin Aggarwal, a machine learning (ML) and artificial intelligence (AI) professional with 9+ years of industry experience in Deep Learning, Data Science, Business Analytics, Management Consulting, Product Management and Software development. Nitin is currently working as Technical Program Manager with Google Cloud, where he leads teams that design AI solutions/products for Google’s strategic enterprise customers.
AIM: How did your machine learning journey begin?
Nitin: I did my MBA from IIT Kanpur and my BTech in Electronics and Communications Engineering from CITM Faridabad. I recently finished executive education at the Graduate School of Business, Stanford University. My fascination with algorithms started from engineering school when I learnt about computer architectures. It got pushed more when I joined my first job as Software Engineer at Compro Technologies, and my mentor Saurabh Seth introduced me to the concept of big data. I also got a chance to work on building some basic data models in excel sheets. I liked it and started exploring this field in detail (primarily Big Data). I started learning statistics, R and SAS at that time. When I joined my MBA, I focused primarily on analytics, statistics and machine learning. My final project was to identify the impact of various vehicle features on sales using panel data. This curiosity grew over time; my career started moving from statistically-based models to AI/ML. When I started working on TensorFlow in 2016 at Google, it gave me real exposure for solving problems at scale, where I had to deal with millions of rows or images. This helped me make a transition from Software engineering to Statistical modelling to AI/ML.
AIM: Would you elaborate upon the initial challenges, and how you addressed them?
Nitin: There were lots of challenges when I started my journey. I didn’t come from a premier engineering college. Lots of opportunities were not open to me. I also faced challenges from getting quality study material and getting the right exposure to solving problems. To address them, I took lots of courses on AI/ML on eDX and Coursera. I solved AI/ML problems on Kaggle and CrowdAnalytix. I used to compete on TopCoder to build my software engineering skills. I connected to so many people on LinkedIn to get practical insights and feedback on my approach. Over a time period, I found that it’s your efforts and persistence that matter. Your college tier and branding don’t matter a lot.
AIM: How do you approach any ML problem?
Nitin: My team is responsible for designing AI/ML strategy for our customers and then executing it by building products/solutions for them. Identifying and formulating the right ML problem is critical for us. We usually start with what problems business is facing by asking questions:
- How are the customers currently solving this problem?
- Can AI/ML be used? and,
- What will other solutions be outside of AI/ML?
This will help us to identify the right problem and the right tools/frameworks. There are so many challenges we face while deploying an ML system. For instance, we recently worked with Seagate to find a way to predict frequent HDD (Hard Disk Drive) problems. Together, we developed a machine learning (ML) system, built on top of Google Cloud, to forecast the probability of a recurring failing disk — a disk that fails or has experienced three or more problems in 30 days. Few top challenges were to get the right data (very low positive rate), define the right business cost for false positive and false negative (to optimise the model performance), automatically identify the model/data drift over a time period, maintain the least downtime (99.999% availability), etc. We used end-to-end serverless Google technologies to overcome lots of these challenges, but it was an amazing ride.
AIM: Tell us what your machine learning toolkit looks like.
Nitin: I usually prefer to build models with existing technologies like BigQuery ML/AutoML Tables for models on structured data, AutoML Vision/NLP for unstructured data, etc. I prefer to use SOTA models instead of going for something brand new. I love using Keras on TensorFlow. I’m a big fan of statistically-based models as well. Hence, I also explore options to build panel data or survival-based models. As we primarily work on Google Cloud, I use Vertex AI Notebooks for prototyping, BigQuery for Data pipelining, and Vertex AI Pipelines for end-to-end orchestration.
AIM: What is a typical day like in your job as a Program Manager at Google?
Nitin: It’s a tough question. My typical week’s work includes syncing up with my engineering team (that includes AI engineers, Data Engineers, Infra engineers, architects and program managers) and discussing/brainstorming our results/risks/ideas on solving the problem. We build the plan on what has been done and what to do next based on the results. Then, we meet with our customers to discuss our findings, brainstorm the ideas and refine our plan. This collaboration will go on and on. I sync up with customer execs to keep them updated and discuss our strategy. I also get involved in the project to do data analysis, building data prep pipelines and build models to get a sense of the data/problem. Along with that, I need to connect with executives from other customers to discuss their AI/ML transformation strategy. I need to collaborate with the AI ethics team, account/sales team, legal, marketing and other product teams. This is just a snapshot, and a lot many ad-hoc things come that I need to take care of.
AIM: What are the various skills you look for in your team?
Nitin: As a Technical PM, I lead teams to build products/solutions for our customers. We are responsible for driving AI/ML transformation for our customers. We work closely with our customers to understand the business problem, formulate it into a technical problem, build a charter and program around it, develop the solution, and help customers launch it as their product. In our team, we have two segments: AI engineering and PMs. For engineers, we lookout for people who can formulate a technical problem, build an AI/ML model and solve it at scale, be curious and learn things whenever needed, be creative and think out of the box. For PMs, we lookout for people who can work in a chaotic environment and bring structure to it. PMs are the face of our solutions. Hence, they must know how to handle various stakeholders like CXOs, engineers, marketing, legal etc. A PM is responsible for maintaining the momentum and bringing the team together. All these skills are paramount to do well in our team.
AIM: What would your advice be to aspirants who want to crack data science/ML roles?
Nitin: I receive this question a lot. My only advice is “Avoid shortcuts”. You can’t leave statistics, maths and coding behind to move on with AI/ML. In the long run, you will suffer. Also, it’s not feasible to learn all of this in 1-2 months. You won’t need all the concepts from day one. Just keep learning and evolving. To crack ML roles, build your profile. It’s high time aspirants move on from the “Titanic” problem on Kaggle or building sentiment models on tweets. Today, to build credibility, one should be capable of showcasing an end-to-end system that can solve problems. Build your GitHub repo, Kaggle and LinkedIn profile to get the right opportunity.
“I believe AI/ML will be democratized even more, and it’ll just become a BI tool.”
AIM: What are some books and other resources that you used in your ML journey?
Nitin: I preferred taking an online course rather than reading a book. I found it more helpful. I especially liked the Analytics Edge course on eDX, Courses by Deep Learning.ai on Coursera, and Executive data science by Johns Hopkins. My favourite book is “The Elements of Statistical Learning” by Trevor Hastie, Robert Tibshirani and Jerome Friedman. There are so many blogs/books available. It’s on you how much you want to explore and learn. Just Google it and enjoy it!
AIM: Which domain of AI do you think will come out on top in the next 10 years?
Nitin: I believe the basics will still stand. Statistical concepts will always be there. I believe AI/ML will be democratized even more, and it’ll just become a BI tool. The most important skill will be problem formulation/solving and understanding the results generated from the model to make any edits. This is where your creativity, experimental thought process and critical thinking come into the picture.
There is so much happening in every domain. Every organization is trying hard to leverage this amazing technology to solve one or more of their business problems. It’s not a race, though. Once use cases are solved in one domain, they can be used in others. For example, solutions developed for predictive maintenance in manufacturing can also be used in Healthcare, for building recommendation engines for Retail and more. Personally, I believe that AI/ML is going to make a huge impact in the healthcare domain and taking it to the doorsteps of it – making it available to last-mile patients – that’ll be the real win for AI/ML.