~/home

Dewanshu Dewangan

Learning machine learning from first principles, toward physics-informed methods.

I'm a student who likes to learn, build, and explore things simply because I'm curious about them.

Academically, I'm drawn to technology, programming, and machine learning. I enjoy understanding how things work beneath the surface, experimenting with ideas, and turning what I learn into projects that I can actually build and improve. I care about having strong fundamentals, but I also believe some of the best learning happens when you stop following instructions and start figuring things out yourself.

Outside academics, I'm interested in music, creativity, culture, and the little things that make life interesting. Music, in particular, is a big part of how I unwind and experience the world — I'm as likely to spend time discovering an artist or a song as I am to spend an evening working on a project.

I'm still figuring out exactly where I want this journey to take me, and I think that's part of the fun. For now, I'm focused on learning deeply, building things I'm proud of, staying curious, and becoming someone who can take an idea from "what if?" to something real.

~/nowtoday

Building a portfolio is one thing. Understanding what it proves is another.

With the 1 November deadline approaching, I've been working on something more useful than adding lines to my CV: building a portfolio of work I can explain, reproduce, and defend.

Three projects are now in place:

  • ThermalLens — An inverse Physics-Informed Neural Network (PINN) that estimates thermal diffusivity from noisy measurements. It includes classical baselines, comparisons, and an error analysis outlining both the results and their limitations.
  • Monte Carlo Methods in Physics — Simulations covering random walks, radioactive decay, and a toy bump hunt, with results checked against theoretical expectations.
  • txt2perseus — A Python converter that transforms plain-text question banks into Perseus-compatible JSON, developed through my freelance work with Khan Academy.

The most interesting thing I've learned came from ThermalLens.

I went into the project looking for a performance result: could a PINN recover a physical parameter accurately from noisy data?

What caught my attention was the similarity between the PINN and the classical fit. Their residuals—the differences between predicted and observed measurements—were correlated at 0.975 on the same dataset.

Rather than showing a clear advantage for the neural network, this suggested that both methods were encountering a similar source of error. It raised a more interesting question: how much of the limitation comes from the measurements and modelling assumptions, rather than the choice of method?

That correlation alone doesn't establish the cause, but it gives me something concrete to investigate. And that's the part I value most: following a result beyond the number and asking what it actually supports.

My next step is not to start another project. It's to go deeper into the work I've already done.

I'm reviewing both repositories line by line, checking the calculations and assumptions, validating the results, and tightening the documentation. I want to be able to explain not just what works, but why it works, where it breaks down, and what I would investigate next.

The goal for 1 November is straightforward: a CV backed by work I understand well enough to defend—not merely projects I can link to.

There's a difference between producing a result and knowing what that result means. I'm working on closing that gap.

all updates →
~/projects2 / 4
all projects →