Projects
Vision42: vision-language system for satellite imagery
Inter IIT Tech Meet 14.0, ISRO SAC problem statement (GeoNLI). Team project.Rank 6
Handles image captioning, visual question answering and text-driven object grounding on satellite images. These scenes contain rotated objects at arbitrary angles and a wide range of scales, so the pipeline pairs oriented bounding box detection (YOLO-OBB) with a multimodal fusion model and serves it through FastAPI. Models were evaluated on GPU infrastructure, using IoU for detection and BLEU-3 for captioning.
Stack: Python, FastAPI, scikit-learn, YOLO-OBB, multimodal deep learning. Source code
Brain functional connectivity from fMRI
Neuroimaging and graph analysis
An fMRI pipeline in FSL and Nilearn that preprocesses 40 ADNI subjects in 12 steps and extracts 1,128-dimensional functional connectivity features. Mixup augmentation grows the cohort 8×. The classification stage benchmarks L1-SVMs, ensembles and 3D CNNs with Grad-CAM explanations. A NetworkX analysis simulates attacks on network hubs; MCI networks turned out unexpectedly resilient, which the results read as computational evidence of compensatory cortical reorganisation.
Result: 97.5% leave-one-out accuracy (F1 0.976) on the original cohort, 99.4% on the augmented data.
Stack: Python, FSL, Nilearn, PyTorch, NetworkX, scikit-learn, MONAI. Source code
DGCNN on Jetson Nano
Edge AI for 3D point cloud classification. Team project.
An edge-optimised Dynamic Graph CNN for point cloud classification on ModelNet10. A Progressive K-Reduction scheme in custom EdgeConv layers cuts KNN edge scatter operations by 37.5% without losing spatial context. A custom torch.cdist graph builder and static graph reuse work around the Maxwell GPU's limits and remove 75% of the KNN latency.
Result: 93.4% accuracy at 18 ms latency with 118K parameters.
Stack: Python, PyTorch, CUDA. Source code
Lunar surface mapping from Chandrayaan-2 data
Inter IIT Tech Meet 13.0, ISRO problem statement. Team project.Rank 11
A four-stage pipeline that turns Chandrayaan-2 CLASS XRF Level-1 data into high-resolution maps of lunar elemental composition. Moving-average noise filtering and Gaussian peak fitting extract quantitative element abundances. PyTorch parallelisation cut processing time by 40%.
Stack: Python, NumPy, SciPy, Astropy, PyTorch, QGIS. Source code
CauESC: emotionally aware dialogue with reinforcement learning
Research project
A dialogue model that combines COMET commonsense reasoning with persona-conditioned generation. The Persona Attention Loop (PAL) keeps the persona consistent across turns. Training runs in two stages: maximum likelihood on a pointer-generator network, then REINFORCE to optimise dialogue metrics that are not differentiable.
Result: BLEU-4 3.92 and ROUGE-L 16.91.
Stack: PyTorch, Transformers, COMET, REINFORCE. Source code
StudySphere
The Forge Hackathon, Meraz 6.0 (OpenLake and GDG)3rd place
A productivity analytics app built during the hackathon. Its Focus Efficiency Algorithm scores focus from usage behaviour instead of raw time tracked.
Stack: React, Vite. Source code