Scientific ML researcher. I take a measurement problem from some corner of science — a leaf's reflectance spectrum, a cell's discharge curve, a microscope's point spread function — and work out what machine learning can honestly claim about it.
The domain changes every time. What doesn't is the discipline: hold something out, report the negative result, freeze the configuration, and ship code that regenerates every number in the paper from deposited artifacts.
- Currently: Research student, Department of Electrical & Electronics Engineering, SSN College of Engineering, Chennai. Four manuscripts under review, each archived on Zenodo before submission.
- Reach me: ORCID 0009-0006-9041-1103
Six problems, five fields, one method: work out what the measurement can support, then refuse to claim more than that.
| Field | Problem, and what came out of it | Artifact |
|---|---|---|
| Plant biochemistry greenhyperspectra-compact-ssl |
Predicting eight leaf traits from reflectance. 30 trait-specific bands beat all 1,721 (mean R² 0.624 vs 0.538) — and self-supervised pretraining, the fashionable answer, turned out to be complementary rather than superior. Reported with the failure: zero-shot transfer to a new acquisition source collapses, and 10–20% local labels only partly recover it. |  |
| Optical physics operator-mismatch-microscopy |
When a microscope's real optics drift from the ones a reconstruction model assumed, can you still bound the error? Two instruments, four reconstruction methods. Severe aberration raises true error up to fivefold while the data-consistency residual goes down — the diagnostic everyone trusts points the wrong way — and 90% coverage falls to 0.31. The one-sided coverage inequality survives every evaluated cell. |  |
| Remote sensing Graph-RMS |
Finding coherent regions in hyperspectral scenes with no labels and no idea how many classes there are. Sparse reciprocal graph → damped diffusion → robust mean-shift modes → reference-free consolidation, held out on Trento. Reports that Leiden on the same graph is competitive, rather than claiming a universal win. |  |
| Battery reliability battery-moo-recursive-forecasting |
Forecasting capacity and resistance recursively, with architecture search given a matched budget against NSGA-II, NSGA-III and random search. 92.5% fewer parameters at statistically indistinguishable accuracy — stated as a parameter result, not an accuracy breakthrough. 12,600-trial log included. |  |
| Reliability statistics discret-hazard |
Where discrete-time hazard models stop being valid for lithium-ion cells: link choice against interval width, effective sample size spanning two orders of magnitude, first-passage against hazard models. RESULTS.md is generated from results/, so the prose and the numbers cannot drift apart. |
 |
| Survival theory grouped-ph-reproducibility |
Separating the information lost by grouping survival times from the error introduced by reconstructing a Cox baseline afterwards: an exact finite-grid information-loss identity, a small-bin expansion, 34 frozen configurations across eight cohorts. With Sudheesh Kumar Kattumannil. |  |
Every package regenerates its reported tables and figures from deposited artifacts — CPU, minutes, no dataset download, no model retraining. Badges are concept DOIs, so they always resolve to the current version.
| Project | What it is |
|---|---|
| ocr | 🏅 National finalist — MOSIP hackathon, IIT Madras Shaastra 2025. Document OCR, field mapping and fuzzy verification (PaddleOCR + FastAPI) wired into a MOSIP pre-registration flow: document-type-aware extraction, normalised fuzzy matching against the applicant's form, and per-field confidence a reviewer can correct. |
| money_muling_detection | Money-mule detection over transaction graphs — PageRank, betweenness, cycle detection, Louvain communities and 72-hour smurfing windows behind a FastAPI + React app. Live demo. |
Python · PyTorch · NumPy / SciPy · scikit-learn · NetworkX ·
conformal prediction · survival analysis · NSGA-II / NSGA-III ·
FastAPI · TypeScript / React · pytest · LaTeX