Chinmay Raut

Project Research Scientist · Health Informatics and Imaging Lab, IIT Madras

Work Experience

Nov 2025 - Present

  • Designed PS-MAE, a masked autoencoder that replaces random patch masking with peritumoral ring masking and Sobel edge reconstruction targets, encoding the clinical prior that nodal metastasis signal lives in the tissue surrounding the lesion rather than within it. Exceeds SOTA under linear probing and holds that margin on a class-imbalanced out-of-distribution external cohort. First-author manuscript under review at IEEE JBHI.
  • Distilled a frozen ultrasound foundation model into a point-of-care-scale ViT via JEPA latent prediction, with a supervised auxiliary loss so the student inherits diagnostic signal rather than agreement with the teacher — 15× fewer parameters while exceeding teacher performance. In preparation, IEEE TUFFC.
  • Built the lab's representation-quality diagnostic toolkit — eigenspectrum analysis, nuisance-subspace alignment, and quantification of acquisition-parameter entanglement — to separate representations that generalize from those that memorize acquisition settings. Diagnosed collapse in an early JEPA-style variant and redirected the pretraining objective accordingly.
  • Showed that breast ultrasound segmentation models degrade across demographically distinct cohorts (Indian, Polish, Egyptian) for acquisition as well as biological reasons, and designed an adaptive wavelet normalization module to close the gap. Manuscript under review.
  • Trained vision-language models to produce diagnostic chain-of-thought for breast ultrasound, using preference-based RL to place reward on the reasoning trace rather than only the final call — so the stated rationale is one a radiologist can audit, not a post-hoc justification.
  • Own the lab's research computing: GPU server specification and procurement, user provisioning and storage allocation, environment and driver stack maintenance, and administration of the lab's GitHub organization.

Aug 2023 - Oct 2025

  • Analysed large-scale telemetric datasets from 400,000+ connected electric scooters each having 100+ IoT sensors transmitting 800 GB+ data daily, extracting insights on riding style, charging patterns, smart features usage and vehicle utilisation across various demographic groups in India, Sri Lanka and Nepal.
  • Built an ETL pipeline using PySpark and Clickhouse SQL for product analytics. Designed and monitored product metrics; conducted regular analysis on their movement to generate actionable product and business insights.
  • Performed population modelling for ride assist features, correlating adoption and engagement of a feature to other ride metrics and behavioural patterns to build adaptive intelligent algorithms, which led to 12% improvement in adoption.
  • Developed a LangChain-based LLM tool to categorise user feedback from multiple channels, unifying qualitative data with quantitative metrics.
Research Intern

Samsung Research Institute Bangalore

Dec 2021 - May 2022

  • Developed a CNN model 'LiteAR' for dynamically estimating illumination for mixed reality applications using spherical harmonics for A and M series Samsung smartphones.
  • Developed an augmented reality app by incorporating LiteAR into the open-source rendering engine Filament.
  • Worked on Samsung's proprietary Neural Radiance Fields (NeRF) model utilising inverse rendering to generate 3D views from a few 2D images.