Skip to content

Now

Last updated: February 2026 · Mumbai, India

Research

Training Spekron, a spectral foundation model that learns transferable molecular representations from IR and Raman spectra. Currently pretraining on QM9S (130K computed spectra) using a CNN + D-LinOSS hybrid architecture on 2× RTX 5060 Ti. Writing up results for the companion paper.

Finalizing the spectral identifiability paper — formal proofs that combined IR + Raman uniquely determines molecular structure for 99.9% of organic molecules.

Building

  • SpectraKit — stable at v1.8.0, 699 tests passing. Exploring wavenumber-aware Whittaker smoothing for the next release.
  • ReactorTwin — hybrid neural ODE digital twin for chemical reactors. Building the online adaptation loop.
  • VibeScope — 3D molecular vibration visualizer. Next.js + React Three Fiber + QM9S data.

Learning

  • State space models — D-LinOSS, Mamba-2, and how selective state spaces handle sharp spectral features differently from attention.
  • Optimal transport for domain adaptation — Sinkhorn divergences for calibration transfer between spectrometers.
  • Conformal prediction — uncertainty quantification for spectral retrieval without distributional assumptions.

Reading

  • Gu & Dao (2024) — Mamba: Linear-Time Sequence Modeling with Selective State Spaces
  • Peyré & Cuturi (2019) — Computational Optimal Transport
  • Angelopoulos & Bates (2023) — Conformal Prediction: A Gentle Introduction

This is a /now page. If you have your own site, consider making one too.