Sarthak Bapte👋

Self-supervised denoising for micro-CT scans, +7 NIQE / +35 PIQE over the baseline.

• Designed a self-supervised Noise2Noise framework using U-Net autoencoders to clean micro-CT scans.
• Optimized training for multi-GPU setups with PyTorch Distributed Data Parallel (DDP).
• Achieved +7 NIQE and +35 PIQE improvement, validated via frequency density histograms.

• PyTorch
• TensorFlow
• Python 3
• Streamlit
• Deep Learning

Decentralized Q-learning agents that coordinate under tightly coupled fire-suppression tasks.

• Developed a decentralized coordination strategy for UAVs using Q-learning and difference rewards.
• Solved pathfinding and fire detection tasks under tight coupling constraints.
• Engineered a custom reward function balancing global objectives against individual agent learning.

• Python 3
• Reinforcement Learning
• Multiagent Systems

Multi-modal U-Net for overlapping tumour regions on the BraTS 2023 dataset.

• Trained a multi-modal U-Net to segment overlapping tumour regions in MRI scans.
• Fine-tuned for region-based segmentation on the RSNA-ASNR-MICCAI BraTS 2023 dataset.
• Implemented advanced data augmentation to handle class imbalance in medical imagery.

• PyTorch
• Python 3
• Deep Learning
• Computer Vision

3D reconstruction from 2D video frames using classical photogrammetry.

• Reconstructed 3D models from 2D video frames using photogrammetry principles.
• Performed feature detection and matching with SIFT and SURF.
• Generated detailed point clouds from matched features using OpenCV.

• Python 3
• OpenCV
• Computer Vision

Virtual PMD camera used to tune intrinsic parameters, with CUDA ray tracing.

• Simulated a PMD camera in a virtual environment to tune intrinsic parameters.
• Developed modules for power calculation and raw data generation using correlation matrices.
• Implemented ray tracing to model light propagation and phase changes accurately.

• C++
• CUDA
• Simulation

Google Daydream VR game, showcased at Unite India 2019.

• Built a mobile VR game with immersive handheld controls using Google Daydream.
• Leveraged C# interfaces and event handling for responsive gameplay mechanics.
• Selected to showcase at the Unite India 2019 Unity developer conference in Kochi.

• Unity
• C#

RTS framework built in 42 days: economy loops, FSM-driven NPCs, A* pathfinding.

• Built a full RTS framework in Unity 3D and C#, including a factory/refinery economy loop.
• Drove autonomous NPC behaviour with finite state machines and A* pathfinding.
• Modelled industrial structures in Autodesk Maya and wired up localized multiplayer.

• Unity
• Autodesk Maya
• C#
• Simulation