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 tumour images in the second dataset that gets used for "external validation" turned out to be duplicates of the first. And under the heaviest ImageNet-C-style noise, the same architecture trained on the source data fell to 38.5% while its calibration error climbed, which is how a deployed model fails without anyone noticing.
That habit 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 a phone assistant over 3,356 government welfare schemes, it meant the model may only state a scheme, amount or rule that a retrieval tool returned, so it doesn't invent an eligibility rule for someone who is counting on the money.
what the chart above is actually measuring
One architecture, four evaluations, each more honest than the last — from neuroscan-ood.
| Condition | Accuracy | What changed |
|---|---|---|
| Leaky split | 96.0% | Random split — slices from one patient land on both sides |
| Grouped split | 90.2% | Split by patient, so no patient is in both sets |
| Corrupted | 66.4% | Mean at the strongest severity of six controlled, scanner-style corruptions |
| Noise | 38.5% | Gaussian noise, the worst of the six |
3,064 slices from 233 patients. The leakage gap holds over ten seeds (paired t-test, p < 0.001); every run is seeded and reproducible.
Research
| Project | What it is | Receipt |
|---|---|---|
| neuroscan-ood | An audit of a medical-imaging benchmark: patient leakage, cross-dataset duplication, corruption robustness, calibration, and label-free fixes. Config-driven package, unit-tested, CI on every push. | 5.8 pts of leakage · 2,335 duplicated images, every pair committed · AdaBN recovers most of the corruption loss with no labels (79.0% → 88.9% at mid severity) |
| Flood-Prediction | Five regressors on 28 years of Kanto-region flood records, tested on a chronological holdout, five rolling-origin windows, and water systems the models never saw. | Accepted at PAUL 2026 · the same ranking under all three tests · paper below |
Built
| Project | What it is | Receipt |
|---|---|---|
| Sarkari-Sahayak | A voice assistant over 3,356 welfare schemes that works on any phone: call a number, press 1 for Hindi or 2 for English, and talk. Built in a team of four. | Top 4, Agent{a}thon 2026 · hybrid dense + BM25 retrieval with reranking · answers only from what its six retrieval tools return |
| MeraPaisa | A native Android IOU and expense tracker, shipped solo. Multi-currency balances, lock-on-edit splits, and rollback to any past entry, with no account and no server. | Kotlin · Jetpack Compose · 636 tests, including every database migration |
| lang-helper-bot | A correction and translation bot, live on Telegram, WhatsApp and Discord: type in Hindi, Hinglish or English, or send an English voice note. | One platform-agnostic engine; each app is a thin adapter, so a fourth platform is one file, not a rewrite |
Also here: NeuroScan-AI — the EfficientNet-B0 classifier that neuroscan-ood later took apart. explainable-vit-chest-xray — a Vision Transformer for chest X-rays, plus a quantitative test of whether its saliency maps can be trusted. MultiClass-Digit-Classification — a 784–64–10 network with every gradient derived by hand in NumPy: 97.26% on MNIST, and 2,000 epochs down to 30 after rewriting the numerical core.
A Head-to-Head Study of Ensemble and Deep Learning Algorithms for Flood Damage Prediction in Japan
P. K. Goswami, J. Arora and Indu · accepted at the 9th International Conference on Pattern Analysis, Understanding and Learning (PAUL 2026) · earlier preprint: 10.5281/zenodo.20084689
Random Forest, XGBoost, SVR, a deep neural network and linear regression, predicting per-capita flood damage across 5,704 events in 75 Kanto-region water systems from 1993–2020. No model is tested on a period overlapping its training data.
Best holdout result: XGBoost, RMSE 0.805 · MAE 0.634 · R² 0.368, with Random Forest 0.0025 RMSE behind. That gap isn't significant (Diebold–Mariano p = 0.45), so neither ensemble is the winner. What holds is the gap below them: the DNN, SVR and linear regression all trail XGBoost significantly (p < 0.001), and the ranking is the same on the chronological holdout, across the rolling windows, and on water systems held out entirely. Catchment area outranks every rainfall variable as a predictor.
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 · Weaviate · Qdrant · LiveKit · 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-oodaudit toward external validation across acquisition sources. - NAT-Test 5Q held, working toward N4: basic conversational Japanese.
- Open to work in Delhi NCR or remote: applied ML, research engineering, or C++/Python systems.
parthkrishangoswami@gmail.com · LinkedIn · Hashnode

