I study Applied AI and Data Science at IIT Jodhpur. Most of what I build turns out to be the same problem wearing different clothes: there is a pile of noisy signal, and something real is buried in it.

Lately that has meant X-ray telemetry from a solar observatory, and Android install traffic from banking scams.

Solar Sentinel  ·  flare nowcasting and forecasting from ISRO's Aditya-L1

Fuses soft X-ray (SoLEXS) and hard X-ray (HEL1OS) across 697 overlapping days, roughly 100 million detector-seconds. Every published nowcaster works from a single soft channel; using both is what buys the extra warning time. Produces a 7,558-flare catalogue at 97.6% recall on C-class and above, with median lead times of 4.8 / 12.8 / 18.1 minutes for C / M / X flares against the published HOPE benchmark's 3.5 / 5.4 / 9.4. The README states the limitations as plainly as the results, because in this field the failure mode is overclaiming.

Shortlisted for the Grand Finale of the ISRO Bharatiya Antariksh Hackathon 2026 at NRSC Hyderabad.

COTTER  ·  fraud detection for Indian bank customers

A scammer can copy an app's name, icon and package ID. They cannot copy its signing certificate. COTTER anchors detection on that, then backs it with permission analysis and India-specific heuristics, and scores APKs, URLs, SMS, screenshots, phone numbers and USSD requests from 0 to 100. One FastAPI backend serves four clients, including an offline USSD channel so a feature phone with no data plan gets the same verdict. Tested against real trojans rather than simulations.

Consolation prize in the final round of the SBI FINNOVATION Hackathon 2026 at IIT Jodhpur, top 12 of the cohort.

Python, SQL, JavaScript. LightGBM, XGBoost, scikit-learn, Pandas, NumPy, SciPy for the modelling. FastAPI and React Native when something has to ship. Matplotlib and Plotly when a chart has to carry the argument.

Jodhpur, India  ·  b24bs1050@iitj.ac.in