The Hometown Bet part 2
The Pivot:
Just like that, the rules of engagement changed. While the candidates before me were graciously allowed to dictate the direction of their interviews by picking a favorite subject, I was greeted by a professor on the other end of the panel with a highly specific question: “Have you done your final year project, or anything else, related to healthcare?”
Internally, I took a sharp breath. I was an Electrical Engineering student. My final year academic project back at NIT Silchar was miles away from anything resembling clinical healthcare. I had walked into that hall armed to the teeth, perfectly prepared to defend complex Machine Learning theory and the rigorous Linear Algebra concepts I had spent months mastering. I was completely thrown off.
But I couldn't backtrack now. I explained that while my core academic project was traditional EE, my true passion lay in the personal projects and internship work I had done in the AI healthcare domain.
That was all the opening the first professor needed. I found myself standing before the panel, diving deep into the architecture of my two core projects: Heartify and Neurafit. As I walked them through Heartify's predictive capabilities using FastAPI and Neurafit's real time tracking utilizing PoseNet, the professor initiated a relentless, rapid fire technical autopsy. I had to defend everything. We went through the datasets, the granular details of my data cleaning processes, and my rationale for the specific models I deployed. They probed every evaluation parameter: How exactly did you test it? What were your accuracy metrics? What specific tech stack and libraries did you rely on?
The Whiteboard Trial: The Second Professor Steps In
During my defense with the first professor, I had mentioned implementing nine different binary classification models. The second professor immediately seized on this detail, and the theoretical grilling began.
"Let's talk about binary classification," he prompted. After I provided the formal definition, he instantly pivoted to K-Nearest Neighbors (KNN).
This was when the solitary whiteboard beside me finally came into play. He instructed me to pick up the marker, draw a 2D dataset on the board, and visually explain the KNN algorithm.
“What would be the classification for K=3? What about K=5?”
I mapped out the data points on the board, drew the boundaries, and gave him the correct classifications. When he asked how I arrived at those specific results, I explained the mechanics of calculating the Euclidean distance.
“Is there any other way of finding the distance?” "Manhattan distance," I replied. "And when exactly would we use Manhattan distance over Euclidean? Which is better, and why?"
I stumbled a bit here. I tried to navigate the answer by explaining that it inherently depends on the use cases such as the dimensionality of the data or whether we are dealing with grid like pathing but I knew I wasn't able to answer that as solidly as I would have liked.
Not letting me dwell on the hesitation, he immediately pressed on: “What is the actual meaning of K? What happens to the model if I increase or decrease the value of K?” I confidently explained the mechanics of 'K' and how it dictates the bias variance tradeoff, finally catching my breath.
The Gauntlet Continues I placed the marker back on the desk, feeling a momentary sense of relief as the second professor nodded and leaned back. But the trial was far from over. The 3rd professor, the 4th professor, and the Head of the Department were up next...