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: Final-year B.E. in Electrical & Electronics Engineering at SSN College of Engineering, Chennai, and Diploma in Data Science and Applications at IIT Madras. Graduating 2027. Several manuscripts under review.
- Reach me: simhaatt2006@gmail.com · ORCID 0009-0006-9041-1103
| Field | Problem → result | Artifact |
|---|---|---|
| Plant biochemistry greenhyperspectra-compact-ssl |
Eight leaf traits from reflectance. 30 trait-specific bands beat all 1,721 (mean R² 0.624 vs 0.538); zero-shot transfer to a new source collapses. |  |
| Optical physics operator-mismatch-microscopy |
Bounding reconstruction error when a microscope's optics drift from the model. Severe aberration raises true error up to fivefold while the residual goes down. |  |
| Remote sensing Graph-RMS |
Unsupervised region discovery in hyperspectral scenes, with no class count given. Held out on Trento; Leiden on the same graph is reported as competitive. |  |
| Battery reliability battery-moo-recursive-forecasting |
Recursive capacity and resistance forecasting under matched-budget NSGA-II/III search. 92.5% fewer parameters at statistically indistinguishable accuracy. |  |
| Battery health sparse-checkup-capacity-forecasting |
Capacity forecasts from sparse checkups: a square-root fade prior plus a gradient-boosted residual, tested on 308 cells and moved to new cell formats with no target labels. |  |
| EV field data DB-PatchKAN |
Driver-aware battery stress and capacity forecasting on 191 real vehicles. Persistence-anchored KAN cuts MAE by 0.009 Ah over persistence (post-hoc). With A. B. Raj and K. V. Shihabudheen. | DOI pending |
| Reliability statistics discret-hazard |
Where discrete-time hazard models stop being valid for lithium-ion cells: link choice, interval width, sample size. Results report generated from the data. |  |
| Survival theory grouped-ph-reproducibility |
Separating information lost by grouping survival times from Cox baseline reconstruction error. Exact finite-grid identity; 34 frozen configurations. With S. K. Kattumannil. |  |
| Survival modelling nestkan |
A discrete-time hazard model that nests grouped Cox, an additive spline model and a gated KAN surface. Every benchmark stage pre-registered. With S. K. Kattumannil. |  |
Archived packages regenerate their tables and figures from deposited artifacts on a CPU in minutes. Badges are concept DOIs, so they resolve to the current version.
| Project | What it is |
|---|---|
| SONA | Final-year project, Phase I complete. Offline speech noise removal: a 722k-parameter GTCRN, about 10× faster than real time on one CPU core, +9.93 dB SI-SDR on VoiceBank-DEMAND. Phase II moves it onto an embedded device. |
| ocr | 🏅 National finalist, MOSIP hackathon, IIT Madras Shaastra 2025. Document OCR and fuzzy field verification (PaddleOCR + FastAPI) for MOSIP pre-registration. |
| money_muling_detection | Money-mule detection over transaction graphs (PageRank, cycles, Louvain, smurfing windows), FastAPI + React. Live demo. |
Python · PyTorch · NumPy / SciPy · scikit-learn · NetworkX ·
Kolmogorov–Arnold networks · conformal prediction · survival analysis ·
gradient boosting · NSGA-II / NSGA-III · ONNX ·
FastAPI · TypeScript / React · pytest · LaTeX