Rohan Pattnaik

SpecPT — Transformer Foundation Model for Spectroscopy

Designed a transformer autoencoder for self-supervised spectral representation learning, trained on 13 million galaxy spectra. The model denoises high-dimensional sequential inputs and predicts continuous targets with R² = 0.99 — compressing a months-long manual analysis pipeline to seconds per sample. Enables zero-shot transfer to new instruments.

Most remarkably: a University of Maryland team adopted SpecPT as the backbone for a mass spectrometry classifier targeting biosignature detection on planetary rovers — with minimal fine-tuning. The representations learned from galaxies transferred to biology.

Redshift Wrangler — Human-in-the-Loop ML at Scale

Built a complete citizen science data pipeline: converted raw FITS astronomical spectral files into visual formats suitable for non-expert annotation, launched the platform on Zooniverse, and recruited 5,000+ active volunteers who contributed 190,000 classifications. The resulting labeled dataset powers downstream ML pipelines with trustworthy training data — achieved at a fraction of the cost of expert labeling.

Cross-Validation Audit Framework for Regulatory Deposit Forecasting

At JPMorgan Chase, designed an audit framework for X-13 ARIMA time-series models used in regulatory deposit forecasting. Identified systematic overfitting in complex model configurations that masked true accuracy. Recommended and validated simpler, interpretable alternatives — improving out-of-sample forecast accuracy by 15% while reducing model complexity and regulatory compliance overhead.

Expert-Labeling Web App + Fine-tuned ResNet (Schlumberger-Doll Research)

Built a containerized expert-labeling web application (Docker + Flask) to replace heuristic-generated labels for rock particle classification. Fine-tuned ResNet on the new expert labels, achieving 85%+ classification accuracy versus the 60% average from heuristics. Replaced a manual, inconsistent process with a reproducible MLOps pipeline.

Real-Time Astronomical Transient Classifier

Built a CNN classifier for real-time detection of transient astronomical events — gamma-ray bursts, supernovae, flare stars — in the Deeper, Wider, Faster survey. Reduced manual inspection by 95%, achieving a 21× speedup in event detection pipeline throughput.