Piyush Malpure

Own the brain of the system. Leading the development of scalable, production-grade autonomy systems for drones. I own the full robotics software architecture and guide the team in building real-time control, perception and autonomy features deployed nationwide. • Spearheading re-architecture from a monolithic, multi-threaded system to a modular, multi-process design using IPC, HAL following design patterns and SOLID principles reducing inter-module dependencies by 75% • Implemented real-time edge-based object detection (<30ms latency) to enable AI-assisted autonomy on embedded drone platforms • Overhauled CI/CD pipeline using GitHub Actions and AWS, reducing release time and deployment complexity by over 40% • Lead and mentor a 6-member autonomy team, shaping architecture, reviewing code, unblocking development, and delivering stable releases • Maintain stability across 60+ customer deployments with ~30 daily flights and <2 support incidents per week • Resolve production-critical and legacy bugs across autonomy, vision, and control systems to ensure field reliability

Worked on core autonomy features for drone navigation, safety, and landing. • Designed and deployed a safety-critical geofencing system with real-time onboard path planning and obstacle avoidance, preventing violations in both autonomous and manual modes • Engineered violation-prevention logic to enforce dynamic no-fly zones, rolled out to 100% of customer deployments • Enhanced vision-based precision landing using ArUco markers and sensor fusion, achieving sub-3cm landing accuracy • Contributed to debugging and stabilizing legacy modules in the autonomy pipeline, improving mission safety and performance

Improved motion planning systems and infrastructure for a 3D warehouse robotics platform, focusing on production readiness, reliability, and IP support. • Enhanced and integrated a custom A* planner into the multi-robot planning stack, optimizing performance and maintainability for dense 3D warehouse environments • Delivered an adaptive contingency service to handle execution failures in real-time, improving system resilience by 2× • Developed a Python ORM layer using Marshmallow and DAO patterns, reducing backend API query latency by 90 seconds • Collaborated cross-functionally to integrate planning logic into scalable production systems • Contributed to patent filings by proposing implementation-specific extensions that helped secure product IP

Developed and evaluated intelligent motion planning strategies for a 3D multi-robot warehouse system, combining classical and reinforcement learning based approaches. • Designed custom 3D path planning algorithms using A* and min-mov logic tailored to dense warehouse environments with high robot traffic • Built and trained reinforcement learning models using Deep Q-Learning and PPO to generalize motion plans and handle edge-case navigation scenarios • Simulated multi-robot environments and benchmarked learning-based planners against classical baselines • Implemented parallelization logic to optimize multi-robot motion planning, improving simulated throughput by 5× • Contributed to research direction and experimental design, forming the foundation for future production deployment

Teaching assistant under Dr. Agheli. Created ROS tutorials for students, guided and cleared doubts, checked assignments, clarified core concepts. RBE 500 - Foundations of Robotics RBE 501 - Robot Dynamics, Graduate Course. RBE 4815 - Industrial Robotics, Undergraduate Course.

Researched and developed deep learning models for radiogenomic classification of brain tumors using multi-modal MRI scans. Focused on mutation prediction, interpretability, and data-efficient learning. • Developed a novel semi-supervised, multi-modal CNN architecture for mutation prediction in brain tumors, improving accuracy by 6% over state-of-the-art benchmarks • Conducted feature engineering and classification using XGBoost, Random Forest, and SVM for radiogenomic modeling • Used class activation maps (CAMs), ablation studies, and visualization techniques to interpret deep learning outputs and validate medical relevance • Integrated multiple MRI modalities (T1, T2, FLAIR) into a unified model pipeline for improved generalization and predictive performance • Contributed to peer-reviewed publications and collaborated with clinical experts to validate outcomes

Built and deployed real-time video analytics pipelines on edge devices using deep learning and streaming frameworks. Focused on computer vision inference, system optimization, and real-time data flow. • Developed an end-to-end video analytics pipeline using OpenCV, GStreamer, and Apache Kafka, deployed on NVIDIA Jetson Xavier and TX2 • Integrated NVIDIA DeepStream SDK for accelerated inference and edge deployment • Trained and tested deep learning models for Pose Estimation and Pix2Pix GAN-based image synthesis, embedding them into production pipelines • Containerized components using Docker to streamline development and on-device deployment • Optimized real-time data flow and GPU utilization to support low-latency vision processing in embedded environments