Sparsh Bhogavilli Email & Phone Number

Tuscaloosa, Al, Us

New York City, New York, Us

- Edenic Energy Inc. (the start-up I am currently working at) is a part of the 2024 cohort of the TechStars Alabama EnergyTech Accelerator in Birmingham, Alabama. - Kindly look at the next tile in the section for a description of my work at Edenic Energy.

- Responsible for all things Data & Machine Learning as a Founding Engineer in the early-stage Climate Tech startup focused on Building Decarbonization.- I also went beyond my job responsibilities to help the start-up with back-end & front-end development in the React, Node.js and Tailwind CSS stack, Data Collection including making phone calls to offices.

Gaithersburg, Md, Us

ML for Modelling & Controlling HVAC Systems:- Working with the Intelligent Building Agents Lab at NIST on reducing the Energy consumption of HVAC equipment in buildings.- Used ML techniques to model the Outdoor Air Emulator - which generates repeatable weather conditions for the HVAC unit.- Developed a Reinforcement Learning based controller capable of controlling the Outdoor Air Emulator. Currently working to make it a practically feasible one.

Gaithersburg, Md, Us

Robotics Simulation platform:- Developing the simulation platform for the Agile Robotics for Industrial Automation Competition (ARIAC) in the (new) Gazebo Harmonic & ROS2.- Enhanced the previous version of the simulator by adding new features.- Ported the existing functionality from Gazebo-Classic to the (new) Gazebo Harmonic.

College Park, Md, Us

- In this project, an observer robot must actively track an adversarial robot with noisy sensors while having no knowledge about the adversary’s actions (Decision Making Under Uncertainty).- Employing an MLP to estimate the noise in sensor measurements.- Currently implementing a path planner based on Bayesian filtering and tree-search techniques.- The work is in preparation to be submitted to Robotics & Automation Letters (RA-L).- Was awarded the Pathway to the PhD scholarship by the Maryland Robotics Center to support this research.

College Park, Md, Us

- Here, a ground rover with no sensors localizes in a known map with a drone's help and navigates to its goal.- Used ArUco markers to estimate the rover's pose from a drone in Gazebo and ran an experiment to find the Pose Estimation error as a function of marker-camera relative pose: - Flew the drone over the marker and recorded the flight in a Rosbag and played it back to calculate 1) actual pose using frame transformations and 2) estimated pose using OpenCV. - Noticed that detecting the marker at long distances depends heavily on padding it with white edges.- Used ROS Navigation Stack for the rover's movement and estimated rover's future pose uncertainty using Extended Kalman Filter.- Uploaded the results as an ArXiv preprint. Revised and submitted to IROS 2023.- Was awarded the Pathway to the PhD scholarship to support this research.

Newark, New Jersey, Us

- Worked on path planning for a fleet of AMRs in warehouse scenarios. Our warehouses had narrow aisles but the existing literature worked only for sufficiently wide aisles. So, I used an existing method as a base & modified it for our use-case.- The resultant method was able to produce feasible plans with minimal wait times for the robots.- In a warehouse simulation with 180 aisle locations & 30 AMRs, it resulted in a throughput of 3200 bins/hr (Each bin’s service time was 15sec & each AMR could carry up to 5 bins at once).- (Single-threaded) Planning time for at least 5sec of runtime was 530ms on a Jetson Xavier NX.- Built a python GUI for visualization.- Used ROS2 to carry the plan data from the C++ planner to the python GUI.

Bangalore, Karnataka, In

- The company works on imaging for defense applications and my project was on multi-task learning of Object Detection + Attribute prediction on thermal imagery from the battlefield.- Modified a YOLO’s final layer to include predicting the object’s orientation & trained it with cross-validation.- Converted the trained model from PyTorch -> ONNX -> TensorRT (on Jetson TX2).- Achieved an inference with 65% mAP (Object Detection) & 0.08 Huber Loss (Attribute Prediction) @ 21 fps.

San Ramon, Ca, Us

- I worked on a Deep Learning for Computer Vision application (Object Detection) in warehouse scenarios & deployed the solution on Edge devices.- Created a custom dataset and cleaned duplicate images using image hashing and irrelevant images using pre-trained object detectors and oversaw the dataset’s labeling process.- Trained a Convolutional Neural Net on this dataset and converted the model to ONNX format & finally to a TensorRT engine on Jetson Nano.- Integrated the TensorRT engine with Nvidia's DeepStream SDK to run inference on a webcam's video feed.- Developed a C++ plugin for DeepStream to post-process the Neural Network's output.- Achieved an mAP of 72% @ 11fps on Jetson Nano.- Demoed the result to the client - which they used to pitch investors.

Hyderabad , Telangana, In

- Curated a dataset of RGB videos for Action Recognition.- Performed KNN classification to determine the most suitable network for our dataset: - Selected a few off-the-shelf Action Recognition networks and passed our videos through them to extract features. - Divided the dataset into 80-20 train-test splits & trained separate KNN Classifiers for the features extracted from each network. - Finally, the network which gave the highest accuracy on the test split of the dataset is selected.- Appended a Auto-Encoder (AE) to this CNN to learn a compact latent space representation for the CNN’s features.- Appended an MLP classification head to the AE’s latent space vector in parallel to the decoder.- Froze the pre-trained CNN and trained the AE using Reconstruction Loss and the Classification head using Cross Entropy and KL-Divergence Loss while employing mini-batch sampling.- Disabled the AE’s decoder during the testing phase and ensembled the predictions across multiple data augmentations to deal with data imbalance. - Achieved a median F1 score of 80%.

Santa Clara, California, Us

- I was part of the Floor Planning team in the Physical Design group.- Worked on improving the quality of assessing the die area utilized by interconnecting metal layers in a chip.- Developed various scripts to calculate the run time of die area utilization computations.- Implemented various checker scripts to validate the automated flows of Floor planning.

Mumbai, Maharashtra, In

- I did a study of various instruments (like Load Cells, controlled valves, etc) used in the Steel Melting Shop - 2 of JSW Steel's Steel plant in Toranagallu, Bellary, Karnataka.- I also spent time in their control rooms studying the PLC Controllers & the ladder logic that automates the processes of their Steel Melting Shop.

Master'S, Robotics

Bachelor Of Engineering, Electronics & Instrumentation Engineering