Some people write about themselves. I'd rather let the mathematics do it.

I didn't start with machine learning. I started with a question — why do systems fail when the data gets harder?

That question led me to IIIT Agartala (B.Tech CSE, 2027), then to a research internship at IIIT Allahabad, and eventually to a stage at the Springer MICA 2025 International Conference where I presented my first research paper.

Along the way, I became a Hackathon Finalist at IIT Kharagpur and the University of Hyderabad.

But more than credentials — I became someone who reads loss curves the way others read stories. Every dip is a chapter. Every plateau is a problem worth solving.

At IIIT Allahabad, I worked on a problem that looked simple on paper but wasn't: predicting heart disease from clinical data where the positive class barely existed.

Most models would learn to ignore it entirely. High accuracy. Zero usefulness. That's the quiet deception of imbalanced learning —

$$\mathbb{E}\left[\left(\hat{f}(x)-f(x)\right)^2\right] =\ \underbrace{\text{Bias}^2[\hat{f}]}_{\text{ignoring the minority}} +\ \underbrace{\text{Var}[\hat{f}]}_{\text{noise sensitivity}} +\ \sigma_\varepsilon^2$$

I refused to accept that tradeoff.

The result was a paper — "Stacking Inside Voting: A Hybrid Ensemble for Heart Disease Prediction" — published in Springer Proceedings at MICA 2025. A hybrid architecture that turned a broken $F_1$ score into something meaningful:

$$F_1 =\ \frac{2 \cdot \text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}} :\quad 0.05 \longrightarrow\ 0.93$$

I reason like Bayes — I don't arrive at conclusions, I update toward them. Every new piece of data shifts the posterior:

$$P(\theta \mid \mathcal{D}) = \frac{P(\mathcal{D} \mid \theta) \cdot P(\theta)}{P(\mathcal{D})}$$

I believe most models fail not because they can't fit — but because they fit too well. Memorisation dressed as intelligence. Regularisation is the discipline that separates the two:

$$\mathcal{L}_{\text{reg}} =\ \mathcal{L}_{\text{data}} +\ \lambda_1\left|\theta\right|_1 +\ \lambda_2\left|\theta\right|_2^2$$

And I pay attention to what a model doesn't know as much as what it does. Uncertainty isn't a weakness — it's information:

$$H(X) =\ -\sum_{x\in\mathcal{X}} p(x)\log_2 p(x)$$

A confident wrong answer is worse than an honest uncertain one.

I work across the full depth of the ML pipeline — from raw tensors to deployed APIs, from research notebooks to production containers.

$$\mathcal{S} =\ \mathcal{S}_{\text{foundations}} \oplus\ \mathcal{S}_{\text{deep learning}} \oplus\ \mathcal{S}_{\text{LLM}} \oplus\ \mathcal{S}_{\text{MLOps}}$$

There is one problem in machine learning I keep returning to — the one that sits at the boundary between memory and adaptation.

Catastrophic forgetting.

When a model learns something new, it quietly destroys what it already knew. Train on $\mathcal{T}_2$ and $\mathcal{T}_1$ is gone. That's not intelligence. That's replacement.

$$\mathcal{L}_{\text{forgetting}} =\ \sum_{t=1}^{T}\sum_{k < t}\ \mathbb{E}_{(x,y)\sim\mathcal{T}_k}\left[\ell\left(f_{\theta_t}(x),y\right)\right]$$

This should go to zero. It rarely does.

I'm working on Dynamic Self-Growing Continual Learning (DSGCL) — a graph-based architecture where the model doesn't replace old knowledge. It grows around it.

New tasks add new nodes. Edge gates $\alpha_{m,k}$ decide exactly how much the past informs the present. Reuse scores $\beta_m$ decide what persists.

$$\mathcal{L}_{\text{DSGCL}} =\ \mathcal{L}_{\text{task}} +\ \lambda_1 \underbrace{\left|\alpha\right|_1}_{\text{gate sparsity}} +\ \lambda_2 \underbrace{\sum_{m}(1-\beta_m)\cdot\mathbf{1}[\beta_m < \phi]}_{\text{reuse penalty}}$$

Plasticity and stability. Not a tradeoff. A single objective.

The belief I build everything on:

$$\frac{d\ \text{knowledge}}{d\ \text{time}} >\ 0 \quad \forall\ t$$

I'm actively looking for ML Engineer, Data Scientist, and Applied Scientist roles where research depth meets production reality.

"The model is only as good as its loss function. Choose yours carefully."