Portfolio

Education

• B.E., Computer Science: Birla Institute of Technology and Science, Pilani
• Relevant Coursework: Generative AI, Introduction to LLMs, Natural Language Processing, Deep Learning, Linear Algebra, Probability & Statistics

---

Experience

• 3D Tubular Instance Segmentation in Biomedical Imaging (Sept. 2025 – Present)

  Working with Dr. Donglai Wei on global reasoning on 3D biomedical images, specifically on thin, elongated structures which require global pairwise point affinity learning to reason over ambiguous structures.

• mmWave Radar to Speech Enhancement (Oct. 2025 – Present)

  Working with Prof. Sandeep Joshi to enhance speech from extremely noisy, band-limited mmWave FMCW radar signals using diffusion- and GAN-based methods, like CDiffuSE and CleanMel+Vocos, respectively.

• Software Engineering Intern @ Stripe (May 2025 - Jul. 2025)

  Saved $200K per annum by reducing 700K manual verifications for merchant addresses using LLMs. Conducted prompt engineering experiments with 1,000 user addresses to reduce the LLM’s false rejection rate to <2%. Also created an abstract framework to reduce time to create new LLM verification systems from ∼2 weeks to ∼1 day.

• Computational Photography using Federated Learning (Aug. 2024 – May 2025)

  Research under Prof. Pratik Narang related to augmenting computational photography tasks, specifically denoising, with federated learning techniques to solve the problem of limited noisy-clean image pair data. Worked on training and evaluating established approaches like DnCNN, DBSN, and Blind2Unblind on custom noise models in federated setups.

  

• Crime Detection in Surveillance Videos (Nov. 2023 – May 2024)

  Worked on the task of detecting and classifying 10 crime types in the UCF-Crime dataset using YOLOv3 and DeepSORT. Preprocessed 1000+ videos, then designed, trained, and tuned an LSTM model, achieving an AUC score of 62.5%, approaching state-of-the-art performance.

  

---

Projects

• 3D Semantic Segmentation of EPFL Electron Microscopy Dataset

  Solved 3D segmentation task on the EPFL Electron Microscopy Dataset using Morphological Post-Processing, nnU-Net, and Foundation Models. Produced a strong baseline Dice score of 94.05% with nnU-Net, then experimented with MIS-FM. Pre-trained PCT-Net using Volume Fusion on the unannotated volume, achieving Dice of 99.93% on the pretext task. Analysed and experimented with preprocessing, Focal Dice, schedulers, and postprocessing to tune downstream training, increasing the Dice score from 42% to 93%.

  

• Implementation of Conditional PixelVAE, PixelCNN for Handwritten Digit Generation

  Implemented the PixelCNN, conditional PixelCNN, and PixelVAE architectures. Extended the PixelVAE architecture by creating a conditional PixelVAE, which achieved strong quantitative performance on binarised MNIST dataset with an average NLL of 0.169 nats per pixel.

  

• From-Scratch Neural Learning in NumPy via Adam Optimisation

  Built forward/backward passes with analytical gradients and gradient descent to study neural learning without frameworks. Re-constructed the Adam optimiser from scratch, yielding 100% test accuracy. Also visualised the model’s predictions via decision boundaries.