• Developed and deployed Pavement Symbol Extraction functionality to the Digital Surveyor software via CUDA-accelarated coordinate frame transformations and segmentation masks.
• Implemented unit and CI testing pipelines with GitHub Actions to validate CUDA kernels and vector-field clustering modules, ensuring production-grade reliability.
• Fine-tuned a Vision-Language Model (VLM) to implement a secondary-inference pipeline to classify extracted open-set painted symbols according to user specifications.
• Conducted 70+ controlled ablation experiments with A/B testing on Hungarian Assigner Costs, Loss weights, Model Queries, Multi-Scale Deformable Attention and encoder-decoder expressivity to boost the performance of the production model by 4%.
• Utilised methods from Object-Detection literature to qualitatively capture a DETR-based polyline detection model's uncertainty to expedite downstream Quality Assurance and Quality Control processes, saving company and customer resources.

• Audio-Visual Navigation: Led development of embodied AI agent with multimodal sensing, training online RL policy with novel class-agnostic reward, reducing path length by 21%
• Offline RL for Indoor Robot Navigation: Built simulation pipeline to collect large-scale trajectory datasets, training a Causal Decision Transformer with early multimodal fusion; integrated environment randomization, behavior cloning baselines, and replay buffer curation to improve policy robustness and sample efficiency
• CLIP-Enhanced Scene Graphs: Designed contrastive-learning framework to compute visual-language embeddings, leveraging GNNs to model object-region relationships
• Open Vocabulary Manipulation (NeurIPS 23): Developed active SLAM exploration algorithm conditioned on probabilistic semantic map, improving task success by 60%.
• Volunteered for the Project Synergy initiative by TCS wherein volunteers taught written and spoken English to students in a Bangla-medium government school.

• Autonomous Sanitization Robot: Designed and implemented end-to-end robotic system during COVID-19 to autonomously sanitize indoor spaces, integrating computer vision, Visual-SLAM, and coverage-based navigation
• Sim-to-Real Deployment: Built Gazebo simulation environments for iterative testing, then transferred stack to hardware platform with onboard sensors and sanitization actuators; finished runner-up among 140 teams
• LiDAR SLAM: Evaluated LiDAR odometry and mapping approaches such as LOAM using CARLA simulation and outdoor driving data, analyzing localization accuracy and map consistency
• Depth Estimation: Implemented stereo and monocular depth estimation methods on driving datasets including KITTI and NuScenes, developing a ROS package for multi-view bundle adjustment

• Created educational content including slides and tutorials for NumPy fundamentals and Loss functions (Focal Loss, Chamfer Loss, RLHF)
• Designed a Colab Notebook and made slides to lead a lab on Backprop and Training Convolutional Neural Networks.
• Collaborated with instructional team to revise and update homework assignments for RNNs, GRUs, Transformers, Language Generation, Diffusion models and PEFT
• Conducted over 40 hours of office hours, labs, and hackathon events, providing hands-on instruction and problem-solving support for undergraduate and graduate students

• CS7.503.M21: Mobile Robotics: The most renowned course of IIIT-H. Provides students with a comprehensive toolkit for research at the intersection of Robotics and Computer Vision, covering SLAM, Computer Vision and Planning algorithms
• CS9.434.S22: Music, Mind and Technology: An interdisciplinary course using algorithms and mathematics to explore how music is perceived by individuals and groups. Served as head TA, designing evaluations for over 60 graduate and undergraduate students
• EC5.205.S21: Introduction to Coding Theory: A fascinating subject building on Shannon's Theory of Communication, exploring the mathematical foundations that underpin everyday communication systems