Most of what I build starts as a doubt about a number.

A brain-tumour MRI classifier of mine reported 95.19%. I spent a summer at DRDO's Solid State Physics Laboratory finding out how much of that was real. Patient-level leakage was worth about 5.8 points on its own. Roughly 85% of the images in the second dataset that gets used for "external validation" turned out to be duplicates of the first. Under ImageNet-C-style corruption the same model fell to 38.5% while staying confident — which is how a deployed model fails without anyone noticing.

That habit is the through-line, and it isn't only a research one. In a neural network I wrote from scratch it meant reading the training loop until I understood why it wanted 2,000 epochs, and then not needing them. In an LLM assistant over 3,397 government welfare schemes it meant the model answers by calling a search tool and never from parametric memory, so it cannot invent an eligibility rule for someone who is counting on the money.

_{Also here: NeuroScan-AI — the EfficientNet-B0 classifier that neuroscan-ood later took apart. explainable-vit-chest-xray — a Vision Transformer for chest radiographs, plus a quantitative test of whether its saliency maps can be trusted.}

A Head-to-Head Study of Ensemble and Deep Learning Algorithms for Flood Damage Prediction in Japan P. K. Goswami and J. Arora · Zenodo, 2026 · 10.5281/zenodo.20084689 · under revision, ICDPN 2026

Random Forest, XGBoost, SVR, a DNN and linear regression, benchmarked on Japanese flood-event data from 1993–2020 under chronological holdout plus five rolling-origin windows. A Random Forest + mutual-information feature-selection pipeline surfaced population and catchment area as dominant predictors over raw rainfall.

Best holdout result: XGBoost, RMSE 0.802 · MAE 0.632 · R² 0.372 — but XGBoost and Random Forest finish 0.004 RMSE apart against a window-to-window standard deviation of 0.077, so neither is the winner. The separation that survives is the one between both tree ensembles and the deep model, which holds in every window.

Languages     C++ · C · Python · Kotlin · Java · SQL · Bash
Systems       Linux (primary) · Git · GitHub Actions · gdb · Make · HDF5 · REST
ML            PyTorch · timm · scikit-learn · XGBoost · TensorFlow/Keras
              NumPy · Pandas · Grad-CAM · torchmetrics
Building      FastAPI · Flask · Qdrant · Streamlit · Jetpack Compose
Foundations   DSA · OOP · Operating Systems · DBMS · Networks · Theory of Computation

• B.Tech, Information Technology (minor in AI/ML) at MSIT, GGSIPU — graduating Aug 2027.
• Extending the neuroscan-ood audit toward external validation across acquisition sources.
• Reading for graduate study in Japan. NAT-Test 5Q held, 4Q booked — basic conversational Japanese, no more than that.
• Open to work in Delhi NCR or remote: applied ML, research engineering, or C++/Python systems.