I build end-to-end machine learning systems for space tech and healthcare — the kind that run on real scientific data and get judged on blind hold-outs, not notebook accuracy.

Most of my work lives at the messy end: raw instrument data, brutal class imbalance, and metrics that have to survive an audit.

_{B.TECH AI/ML  ·  RAIT, DY PATIL UNIVERSITY  ·  CLASS OF '27}

_{upper management · GIF by Pixel Jeff}

EOS 6  

_{TASK-ROUTED MIXTURE OF EXPERTS}

Hand it raw solar X-ray telemetry and it does the rest — builds its own features, routes them through every expert, and streams one combined result: what is flaring right now, and the probability of a major flare in the next 24 hours.

LUMEN 6*  

_{P(MAJOR FLARE, NEXT 24H)}

Gradient-boosted ensemble over magnetogram-derived active-region features, with calibrated probabilities.

SPECTRA 6*  

_{LIVE FLARE STATE + CLASS}

Twin temporal convolutional networks reading per-minute X-ray telemetry, one per channel.

ROENTGEN

_{solar flare nowcast + 24h forecast}

Mixture-of-experts over dual-channel solar X-ray telemetry, soft and hard: two TCN nowcasters, a gradient-boosted forecaster, and an independently built flare catalogue. Above the published attention-based SOTA on SWAN-SF.

EINTHOVEN

_{ecg pathology detection + dataset audit}

A 1D ResNet from the PTB-XL benchmark sorting 12-lead ECGs into five diagnostic superclasses, tested on PTB-XL's official patient-stratified test fold (2,158 ECGs) and reproduced exactly from saved predictions. Adding three 3D heart-vector (VCG) channels was a null result: +0.00055 AUC, with a 95% CI that crosses zero.

An Einthoven's-law audit of all 21,799 PTB-XL recordings flags 57 (0.26%) whose limb leads disagree, and PTB-XL's own noise labels don't predict which. The audit CSV is released. The website shows 100 test ECGs on clinical paper with the model's scores and a rotatable 3D heart-vector view.

KENYON

_{a fruit-fly memory circuit built from its own connectome}

The fly's learning centre, the mushroom body, taken from the MaleCNS brain scan (projection neurons → 4,064 Kenyon cells → 97 output neurons), built from the scan's real synapse counts, with the published dopamine learning rule. It learns which odour predicts punishment or reward, holds a punishment and a reward memory at once, and shifts a simulated choice. One learning rate is calibrated to Hige et al. 2015; where the memory lands (96% of it on the paired output cell), how specific it is and which way the choice goes come from the wiring. Negatives reported: no pattern completion, and a second memory in the same compartment overwrites the first.

Does the memory reach the body? Barely: on the full graph, 2 of 1,314 descending neurons move (by 2.8% and 0.91% of their rates), through 9 direct synapses. Cut those and the effect collapses.

NEUROPHARMA

_{drug-drug interaction checker}

Checks every drug pair against DDInter, a curated database of 169,395 graded pairs across 1,610 drugs, shipped inside the page so lookups run in the browser. A pair with no curated grade shows as "Not documented", never as safe. For those gaps, an opt-in LightGBM model on chemistry features can give an estimate (it abstains when unsure), clearly labelled experimental; its accuracy on those pairs is not yet measured.

React + Vite on Vercel, FastAPI in Docker on Render. Research prototype, not a medical device.

MAGNUSON

_{geomagnetic storm interval detection}

A nowcaster that flags storm-driving solar-wind intervals in Aditya-L1 magnetometer data. Bx/By/Bz/|B| readings become 9 physics-informed features, a PatchTransformer (trained on 128-minute windows of one-minute data) scores them, and an HMM/Viterbi pass smooths the scores into clean intervals. It detects; it does not forecast onset. On data never used in training it flags the 10–11 October 2024 storm: one event, run on the sensor's raw 10-second data rather than the one-minute data it was trained on, so a sanity check rather than a benchmark.