About
I'm Harsh. I'm in my final year of a B.Tech in computer science, with a minor in AI/ML, at Sharda University, Greater Noida, finishing in June 2027.
Research
Most of my time goes into two research projects. Both have been submitted to conferences, and both taught me more by failing than they would have by working.
The first is RareSight. It takes a patient's symptoms, written as standardised phenotype codes, and tries to rank which of 557 rare diseases they might have. I built a graph neural network for it, and the graph network lost badly to a much simpler method that just compares symptom lists directly. Rather than swap in the thing that won and move on, I spent the rest of the project working out why, and ended up proving that a model trained on the same ontology it's tested against can't beat that ontology used directly. Along the way I found that the benchmark everyone uses quietly rewards models for repeating the patient's own words back at them. That turned into the paper's main argument.
The second asks whether you can tell when your training code is silently wrong. I broke the backward pass on purpose in one specific layer, then ran 7,070 training runs across five datasets to see whether the damage showed up in how the model behaved. Past about two layers deep, it doesn't. The network absorbs the fault and trains to nearly the same accuracy, which is an unsettling thing to learn about code you'd otherwise trust. All of it ran on one RTX 3050, mostly overnight.
Work
I also spent six weeks interning at Siddhi Sarees, on a team of about a dozen people rebuilding their site for international customers. My actual job there was small: bug fixes, layout problems, the things interns get handed. But nobody could tell where an order was or who was holding it up, so I built them an internal dashboard for it in plain JavaScript with no framework. Nobody asked me to. It's on GitHub at github.com/hersh11/dashboard-siddhisaree.
Tools
On the practical side, it's mostly Python and JavaScript, PyTorch for anything with a model in it, SQL and Java when I need them. I've used SageMaker, S3 and EC2 in our AWS lab, written GitHub Actions workflows and poked at Jenkins. My Linux is functional rather than good. I've used Docker in coursework but wouldn't claim it yet.
Outside code
Away from all that I play a lot of video games, mostly FPS, I like design, and I travel when I get the chance. I volunteer too. Before university I spent a few months with TechBridge India Foundation in Bareilly running digital literacy and beginner coding workshops for school students. There were no lesson plans, so I wrote them. Explaining a for-loop to someone who has never opened a terminal is the fastest way I know to find out whether you understand it yourself.
What's next
I'm looking for a software or ML internship now, and I'm curious about working somewhere with a different engineering culture to the one I've trained in.
The quickest way to reach me is email. My resume is on Google Drive, and everything else is on the links page.