Human decision making, as simplistic as it might sound, relies on a complex framework of observations and perceived outcomes that drive our actions !
Nishant Bhadauria, our next pathbreaker, Applied Mathematician & Statistician at AT&T, applies Reinforcement Learning techniques to train Virtual Assistant Bots.
Nishant talks to Shyam Krishnamurthy from The Interview Portal about machine learning being an ever evolving domain with state of the art algorithms that keep improving day-by-day to address real world challenges.
For students, the diversity of business problems that ML can address is astounding. The challenges are limitless and so are the opportunities !
Nishant, your background?
I grew up in a family of 5. My father is a retired banker and my mother is a homemaker. I have 3 siblings, I being the eldest. We grew up in Saharanpur/Dehradun.
What did you do for graduation/post graduation?
I studied Maths (BSc) in BHU, followed by an MBA in Power Management in UPES. I wasn’t too good in studies, but discovered my interest very late in my life just at the end of my MBA. I discovered the field of Econometrics, which helped me see what statistics can really do.
What made you choose such an offbeat, unconventional and uncommon career?
I was influenced by Geoffrey Hinton (for his work on Deep Learning) and Von Rossum (known as the creator of the Python programming language) a lot.
Though I am not a very adamant follower of influencers, anyone who does good work in statistics, programming etc. inspires me.
I have read a few books: ISLR for R , SAS book by Ron Cody, and ML (Machine Learning) course by Andrew NG.
I think the most significant turning points were learning SAS and SQL during my MBA, which gave me an opportunity to get into the industry.
While working in UHG (UnitedHealth Group), I learned R and Python, and ventured into ML
I have been a contributor in Kaggle where I have learnt the latest techniques and alogs
How did you plan the steps to get into the career you wanted? Or how did you make a transition to a new career?
It was gradual, from SAS reporting to ML. Once I got a good command of statistics, from reporting basic stats in dashboards to building linear and logistic regression models, I realized that ML is always the next step, which is followed by Deep Learning, Reinforcement learning and so on.
During my college days, around 2009 , SAS was still a very rare skill and in high demand. So, I learnt SAS which helped in getting started with my career in Analytics. My approach has always been to move in the direction of the industry and learn new skills with time. Learning R and Python in 2013/14 when the ML boom was beginning helped me immensely. The next goal in the current situation is Reinforcement Learning and Quantum Machine Learning. This will be the future.
All my jobs have had the requirement of a partial or complete ML cycle. Though the domains have been varying and applications have been different, the fundamentals of programming and analysis remain the same.
In TCS, I worked in the Media division of Nielsen Media Research, where I measured viewership and predicted future viewership of television, on the basis of which the broadcasting channels charged advertising rates.
My next role was at UHG (Health insurance) where my goal was to score claims based on probability to turn into an appeal and close them proactively, creating rules of COB’s and finding patterns and models for provider fraud. The data sources were very extensive, from ICD codes to Procedure codes to historical records.
In my next role at Hewlett Packard, I primarily worked on reducing executive escalations and mining server logs using NLP. We employed various clustering techniques on text data, for most frequently occurring problems and troubleshooting done by engineers. We did adhoc analysis to reduce the parts used in repair and maintenance.
In Citi, I worked as a part of the Anti Money Laundering team which helped in scoring alerts generated out of the Transaction Monitoring System. Based on statistical analysis, we did threshold tuning of scenarios and looked at patterns of transactions vs KYC and third party data to identify shell companies.
In my current role at AT&T, I am part of a team that is training a VA bot (Virtual Assistant) that negotiates with customers for payments during a conversation. I work on designing an environment that does Grid search based RL (Reinforcement Learning) to decide the best option among possible options available during the call.
How did you get your first break?
It was through college. HCL visited our college campus looking for people who had knowledge of statistics and a few tools, like Minitab or SPSS. When they came to know I knew SAS , they took my interview based on SAS and basic statistics and I was selected and started working even before my final semester exam.
What were some of the challenges you faced? How did you address them?
Challenge 1: Main challenges are around data collection, extraction and preprocessing, not to forget regulatory and team specific constraints. There are also challenges in getting quality data in time, in the desired format or converting it to the desired format.
Challenge 2: Sometimes nothing works, none of the algorithm give you desired results. In such cases, we have to decide on what works, whether to go with business rules and how to convince management with results.
Challenge 3: Production environments are the biggest challenge, the teams that manage production environments can be very specific and demanding in terms of the kind of results they want and what things can be accommodated and not accommodated. Very basic things such as library version of Python can halt entire production. Hence, these things have to be kept in mind.
Where do you work now? What problems do you solve?
I currently work at AT&T where i am training a VA bot using RL.
Reinforcement Learning is a technique where an agent (software program) interacts with an environment (virtual or real) and takes action according to rewards which are based on policy. Most common examples of reinforcement learning are self driving cars.
What skills are needed in your role? How did you acquire the skills?
I need to have a good grasp of NLP, Python, RL and MLflow.
What’s a typical day like?
We are assigned tasks at the beginning of the cycle. We typically have to develop or deploy a model by a certain date, and the subtasks are to be completed in a day or 2, like adding a variable, checking the hyperparameters, running the results and reporting metrics, or adding more data, or acquiring more data sources etc etc.
What is it you love about this job?
I like the non-repetitive and the challenging nature of work.
How does your work benefit society?
Well, in the corporate environment, we primarily help in making decisions based on data. A data scientist on the whole helps make informed decisions that help in reducing cost, reducing frauds, and optimizing operations that improves customer experience .
Tell us an example of a specific memorable work you did that is very close to you!
There are many. If I have to choose one, it will be identifying faulty/burnt parts of servers using Convolutional Neural Networks (CNN), which we did in Hewlett Packard Enterprise. It helped us quickly identify 10 common failures and map them to solutions, which rescued TAT by whopping 80%.
Your advice to students based on your experience?
1.Be open to learning new things, data science is an ever evolving domain and every year State of the Art algorithms keep coming up or improving. Since new libraries are getting added daily, keep track of useful ones and learn them by all heart
2. Practice a lot. Best platform is Kaggle, which is free, where you have support from the community and Google resources like TPU GPU.
3. Focus on Business problems and mathematics as well. It would be good to have command in Linear Algebra, Calculus and Statistics. First, understand the problem before even writing the first line of code. Break the problem into what could be achieved and how you would approach it. More often than not, business problems do not require sophisticated models. Basic things like pivots will do. Hence, you should be vigilant to these things
- Learning Production side of things like Azure and Kubeflow, Microservices
- Applying more Reinforcement Learning, this is the future.
- Learning Quantum Machine Learning , I am currently learning this using IBM Qiskit and am probably hoping that something might be applicable in near future.