When Harish began thinking about the problem he wanted to work on, it wasn’t abstract. It was practical. In many parts of the world, access to timely medical diagnosis isn’t guaranteed, and delays often have real consequences.

He wasn’t trying to solve the entire system. He was trying to understand where technology could reasonably help.
“I kept coming back to the idea of access,” he says. “Not everyone has the same tools, or the same amount of time.”
That framing shaped everything that followed.

Learning to Work With Imperfect Conditions

Early on, Harish realized that real-world data rarely arrives in clean, ready-to-use form. Images had to be organized, resized, labeled, and handled in batches. None of this was conceptually difficult, but it demanded patience and attention.
“It made me slow down,” he reflects. “If something was off early, it affected everything later.”

Instead of rushing toward results, he began paying more attention to structure. How inputs were prepared. How assumptions were embedded into the process. How small decisions accumulated.
That awareness stayed with him.

Understanding Models Beyond Accuracy

As the work progressed, Harish spent time training and testing a model. Accuracy numbers were encouraging, but they weren’t the point.

What mattered more was understanding why performance changed. Why results differed when the model moved from a controlled environment to a live interface. Why deployment introduced new constraints.

“When it didn’t behave exactly the same, I had to think about reasons,” he explains. “Storage limits, preprocessing, real inputs. Things I hadn’t considered at first.”
That gap between theory and practice became instructive.