I am an undergraduate student in Artificial Intelligence at the Indian Institute of Technology Gandhinagar. Currently, I am a Visiting Undergraduate Researcher at California Institute of Technology (Caltech), advised by Dr. Ashish Mahabal, investigating fine-tuned vision-language models for clinically grounded abdominal 3D CT analysis. Previously, I was a Summer Undergraduate Research Fellow (SURF) at Caltech developing AI-readiness frameworks and Croissant-compliant data pipelines for medical imaging collections.
My research focuses on developing AI systems for medical imaging and visual understanding. I work on computer vision, generative models and vision-language models, with an emphasis on applying these methods to biomedical domain.
- [Aug 2026] Continuing as a Visiting Undergraduate Researcher at Caltech
- [May 2026] Started Caltech SURF fellowship.
- [Dec 2025] Awarded Diffusers MVP by Hugging Face for impactful open-source contributions.
- [Aug 2025] Received 2nd Prize at IITGN Undergraduate Research Showcase for Histotripsy bubble segmentation in ultrasound images.
- [Aug 2023] Started my journey at IIT Gandhinagar as an undergrad.
- Investigating vision-language models for answering clinically grounded questions about abdominal 3D CT scans.
- Developed and validated a quantitative ten-dimension framework to assess AI-readiness of medical datasets, applied it over 10 cancer collections, and identified recurring gaps in label fidelity and operational readiness.
- Developed pipelines transforming multi-modal Early Detection Research Network (EDRN) cancer datasets into Croissant-compliant, AI-ready resources for large-scale machine learning applications.
- Created tutorial notebooks demonstrating exploratory ML workflows on AI-ready medical datasets.
- Investigated whether injecting lightweight semantic (DINOv3) and structural (PiDiNet) guidance directly into a frozen SDXL diffusion prior could eliminate the need for disjoint pre-cleaning networks in blind image restoration, training on 30K+ images across multiple real-world datasets.
- Designed a two-phase training regimen consisting of coarse structural alignment via pixel-space regression at high-noise timesteps, followed by perceptual and semantic-similarity refinement at low-noise timesteps.
- Discovered that continuous low-quality latent injection anchors the diffusion prior to the input, and formulated a Restoration Sampling Strategy (RSS) that thresholds this injection across the sampling trajectory achieving an SSIM of 0.50 under compound degradation.
- Investigated whether modern segmentation architectures and a combined Dice-Focal loss could resolve a known early-stage detection failure in ultrasound-guided histotripsy monitoring, benchmarking four architectures against an established ResNet-18 baseline.
- Introduced a Combined Dice-Focal Loss and a pulse-aware weighted sampler to address class imbalance from sparse early-stage bubble clouds, improving mean IoU from 0.72 to 0.84 over the baseline.
- Found EfficientNet-B0 U-Net best suited for real-time deployment (7.9M parameters, 13.5ms inference) and DeepLabV3 best for temporal tracking fidelity (Pearson R = 0.92), revealing an accuracy-versus-tracking trade-off.
- Contributed to the development of SARTHI, a domain-adapted agricultural advisory LLM fine-tuned from Gemma-3-27B-IT using LoRA Supervised Fine-tuning across 16×H100 GPUs on a 220K-sample synthetic chain-of-thought dataset for taxonomy-driven agronomic reasoning.
- Developed LitGPT-to-HuggingFace conversion tools for standard serving stacks, indexed a 35GB+ continual pretraining corpus, and experimented with prompting techniques for structured reasoning.
- Contributed FLUX.2 Klein inpaint pipeline, resolved critical hardware compatibility and inference bugs across multiple generation pipelines such as FLUX, Kandinsky, and Cosmos.
- Awarded the Diffusers MVP title by Hugging Face for impactful open-source contributions.
- Resolved generation and export bugs related to RoPE KV cache, distributed serving, and checkpoint conversion.
- Engineered a lightweight CLIP model pairing ViT-S/16 and ResNet-101 vision models with a RoBERTa text encoder fine-tuned on Flickr30K dataset to learn joint visual-semantic embeddings.
- Achieved an 83.8% Recall@10 in image-to-text retrieval, matching the zero-shot OpenAI CLIP baseline.
- Engineered a scalable graph-streaming library implementing linear sketches, Disjoint set union-based connectivity, sampling, and randomized algorithms for large-scale graph analytics.
- Implemented and benchmarked Count-Min/Count Sketches and streaming graph algorithms on real-world DBLP and Enron graphs, achieving 2.95K edges/s with Count-Min Sketch at only 0.05 MB peak memory and reducing connectivity runtime by 59% and memory by 81% compared to NetworkX baseline.
- Developed and evaluated streaming optimizations including k-skeleton certificates, cut sparsification, greedy weighted matching, and shortcut-based random walks while processing at 1.2–3.0M edges/s.
- Reproduced the "Pruning via Merging" (MKA) paper, building a pipeline to compress LLMs by merging layers based on manifold alignment and validating the baseline performance at various compression ratios.
- Proposed align-then-merge mechanism that uses the Hungarian method for optimal permutation of neurons to align them before merging, outperforming prior results by +9% accuracy on the MMLU dataset using Llama3-8B.
- Reviewer: NeurIPS 2026 (Ethics Review Track)
- In my free time, I enjoy graphic designing, playing badminton, and mobile photography.
- Check out some of my clicks and designs.
$