Abhiroop Tejomay Kommalapati

Master of Science in Computer Science

Indiana University Bloomington | Bloomington, IN | May 2024 | GPA: 3.89/4.00

Courses: Applied Algorithms, Elements of AI, Deep Learning Systems, Computer Vision, Applied Machine Learning, Software Engineering

Computer Vision Research Assistant

Indiana University Computer Vision Lab | Bloomington, IN | May 2023 - May 2024

• Engineered novel latent diffusion-based depth map refinement and super-resolution networks to enhance depth map quality in multi-view stereo networks, optimizing 3D scene reconstruction outcomes.
• Developed a novel latent diffusion-based network that regularizes intermediate 3D cost volumes from a pre-trained multi-view stereo network to improve depth map estimation and 3D reconstruction.

Machine Learning Researcher

Indiana University Vehicle Autonomy and Intelligence Lab | Bloomington, IN | Oct 2023 - May 2024

Tech Stack: ROS2, YOLOv8, LIDAR, Sensor Fusion

• Pioneered advanced sensor fusion modules in ROS2 integrating Camera, LIDAR, Radar data as part of the perception sub-team of the IU Autonomous Racing Team building high-speed autonomous racing cars that can reach 170mph.
• Enhanced the perception subsystem by developing object detection modules based on YOLOv8 and LIDAR filtering and clustering modules, and implemented sensor fusion techniques for reliable and accurate representation of vehicle environment.

Machine Learning Engineer

AiKYNETIX LLC | Houston, TX | May 2023 - Dec 2023

Tech Stack: PyTorch, CoreML, Docker

• Ported state-of-the-art human pose estimation models from PyTorch to CoreML format for AiKYNETIX's human motion video analytics iPhone application, improving performance of the app by 25%.
• Collaborated with the web development team and orchestrated the seamless launch of AIK-Web, AiKYNETIX's cutting-edge web-based platform for human motion analytics, with a 20% increase in the accuracy of the metrics.
• Deployed the web application end-to-end using Docker with support for switching different pose estimation models.
• Improved video processing pipeline by 30% by caching results of previously processed frames and reusing results for similar new frames.

Computer Vision Engineer

Segmind | Hyderabad, Telangana, India | Apr 2022 - Jul 2022

Tech Stack: PyTorch, ONNX, TorchScript, TensorRT, Apache TVM

• Developed VoltaML, a library to accelerate ML models using target hardware optimization, compilation techniques, and network quantization.
• Accelerated inference by up to 10x compared to native PyTorch by leveraging ONNX, TorchScript, TensorRT, and Apache TVM.
• Dockerized accelerated models and collaborated seamlessly with a 5-member web development team to ensure seamless integration.

Machine Learning Engineer

Onward Health | Hyderabad, Telangana, India | Jan 2020 - May 2022

Tech Stack: OpenCV, PyTorch, U-Net

• Implemented an end-to-end medical image stitching tool using OpenCV that stitches frames captured from a camera-mounted microscope in real-time at 20 frames per second.
• Pioneered U-HoVerNet, a novel image segmentation network to segment cells in whole slide images, scoring 6th place (post-challenge leaderboard) in the MoNuSAC grand challenge and published paper in IEEE TMI.
• Designed object detection models to detect lung nodules in chest X-rays and CT imaging scans achieving an IoU of 76%.
• Conducted interviews for hiring new ML engineers to the team and coached 3 interns within the company who now work full-time.

HospQA

Tech Stack: GPT-3.5, RAG, Neo4j

Built a chatbot utilizing LLMs (GPT 3.5) and Retrieval Augmented Generation (RAG) techniques for context-aware responses to inquiries about hospital systems by retrieving data and running queries on a graph database about the patients, patient experiences, and visits.

Image to Prompts

Tech Stack: BLIP, CLIP, Vision Transformers

Designed an image captioning 4-model ensemble based on BLIP, CLIP and Vision Transformers to reverse engineer the prompts used for image generation models like Stable Diffusion, leading to a better understanding of engineering prompts.