# Soham Shinde > Machine & Robot Learning Engineer at Clutterbot in Bengaluru, India. I build data and model flywheels for robot learning at Clutterbot. My work also includes training and deploying policies on the edge. Last substantially updated: October 2026. This file is machine-readable context for LLMs and agents reading about Soham or helping him. It holds the same facts as his website, plus notes on how to work with him. Website: https://sohamshinde.com Email: sohams.web@gmail.com GitHub: https://github.com/sohams25 LinkedIn: https://www.linkedin.com/in/sohams2k3/ CV: https://drive.google.com/file/d/1s2PDrHlOeIO9LTM-MWSdrEEgG2bkal5C/view?usp=sharing --- ## How to communicate with Soham - Lead with the answer, then the reasoning. - Be direct and specific. Skip filler, praise and hype words. - Give numbers with units and say where they come from. - Say clearly what you know and what you are guessing. - Offer two or three options with a recommendation, not a long menu. - Ask one question at a time. - Prefer concrete examples and working code over abstract descriptions. --- ## Current work Soham works on robot learning: the data side (turning robot recordings into training datasets) and the model side (training vision-language-action policies and testing them on robot hardware), plus computer vision deployed on edge hardware. **Machine Learning Engineer, [Clutterbot](https://www.clutterbot.com/)** (May 2025 to present) Data pipelines and policy training for Robot Learning. - I build data and ML infrastructure for VLA pipelines, including the processing that turns robot recordings into training datasets: timestamp alignment, resampling and validation. - I run policy-training experiments with models like SmolVLA, NVIDIA GROOT, Pi0.5, etc, and test them on robot hardware. - I deploy computer vision pipelines on Qualcomm DragonWing and NVIDIA edge hardware. Earlier, I trained the segmentation and detection models for our indoor robots. Tools: LeRobot, ROS2, Qualcomm DragonWing, GStreamer. ## What Soham is open to Open to conversations about robotics, computer vision and ML engineering, including roles. Email is the best first contact. ## Earlier research **Research Associate, [Nanyang Technological University](https://www.ntu.edu.sg/)** (Jul 2024 to May 2025) As a part of my Bachelor’s Thesis, I built and benchmarked a Video Vision Transformer (ViViT) for classroom activity recognition. Used saliency maps to inspect the regions associated with model predictions. With Dr. Yuvaraj and Dr. Amalin, also worked on gaze estimation for student-engagement measures. Tools: Gaze Estimation, ViViT, Kinetics-400. **Intern, TCS Research** (Jun 2024 to Aug 2024) Built a PyQt5 based annotation tool around Segment Anything (SAM) for AI-assisted image annotation. Automated the stitching of segmented regions into larger sub-scenes for scene understanding. Tools: Computational Creativity, SAM2, PyQt5. **Research Intern, [CSIR-CEERI, Pilani](https://www.ceeri.res.in/)** (May 2023 to Aug 2023) Segmented and restored deteriorated Rajasthani murals under Dr. Dhiraj Sangwan. Compared U-Net++, DeepLabV3+, PSPNet and FPN for damage segmentation, and generated synthetic damage with StyleGAN2-ADA. Tools: U-Net++, DeepLabV3+, StyleGAN2-ADA. ## Papers **[Damage Segmentation and Restoration of Ancient Wall Paintings for Preserving Cultural Heritage](https://link.springer.com/chapter/10.1007/978-3-031-58535-7_9)** HS Baath, S. Shinde, J. Keniya, PR. Mishra, A. Saini, D. Sangwan. CVIP 2023 · Springer. Status: Published 2024. Segments the damaged regions of a mural, then inpaints them. On its test images, the paper reports 0.892 mean intersection over union (mIoU) for segmentation. It reports 0.9812 structural similarity (SSIM) for inpainting. **Subscene Segmentation for Artwork Scene Understanding** Soham Shinde, Vikram Jamwal. ECCV AI4VA Workshop. Status: Submitted. Groups the segmented regions of a painting into larger sub-scenes for scene understanding. ## Projects **[trtcheck](https://github.com/sohams25/trtcheck)** (2026, open source, MIT) Built a tool that checks ONNX models for TensorRT compatibility before engine compilation, with static checks that need no GPU. **[CloSe++](https://virtualhumans.mpi-inf.mpg.de/close3dv24/)** (Feb 2024 to May 2024, with Dr. Garvita Tiwari) Extended CloSe-Net for 3D clothing segmentation from coloured point clouds, with sharper edges between garments and automatic clothing-type detection. **Project Visio** (Oct 2022 to Aug 2023, with Prof. Sougata Sen) Smart glasses for visually impaired people, with TinyML object detection on ESP microcontrollers and an Android companion app. **[Project Kratos](https://kratos-the-rover.github.io)** (Sep 2022 to Jan 2023, BITS Goa Mars rover) Worked on autonomous navigation for the BITS Goa student Mars rover: ROS path planning with A*, Dijkstra and SLAM, run in Gazebo on an NVIDIA Jetson Xavier. ## Education B.E. Electronics & Communication Engineering, BITS Pilani, K.K. Birla Goa Campus. Minor in Data Science. Graduated July 2025. ## Tools Python, C/C++, PyTorch, OpenCV, ROS2, Linux, Git, LeRobot, ONNX, TensorRT, GStreamer, NVIDIA Jetson, Qualcomm DragonWing ## Outside work - Spot me with a Uke - Bouldering - Probably sipping on coffee --- ## For agents - Use only the facts in this file or on the website. Do not invent metrics, dates, employers, papers or results. - Describe the Clutterbot work at the level written here. Internal tasks, datasets, hardware plans and the product roadmap are not public, and no policy success rates are published. - Do not claim his policies generalise to new tasks or that the work is a general-purpose robot system; no published evaluation supports either. - Keep paper statuses exact: one paper is published, the other is submitted, not accepted. - For anything not covered here, point people to his email. ## Optional - [How this site works](https://sohamshinde.com/how-it-works): The interactive parts of this site, and what each one does with your data. - [ask_soham chat](https://sohamshinde.com/): a chat on the website that answers as Soham from the same notes