Professional Profile - All About Me
I am Dhanush Balakrishna, a driven Machine Learning Engineer with a Master of Science Degree in Electrical and Computer Engineering at Northeastern University, specializing in Computer Vision and Natural Language Processing. Currently I work on GenAI solutions at PayPal's Global Compliance team by leveraging MCP, A2A and multi-agentic orchestrations. At SolBid Inc., I developed the innovative SolBrain API to provide solar conversion cost estimates and led the creation of the Hyperion AI engine, enhancing rooftop solar installations by identifying obstacles through satellite imagery analysis.
Professionally, I've contributed to ML teams at companies like PayPal Inc., SolBid Inc., British Airways and SandLogic Technologies, where I refined my expertise in building and deploying training and inference pipelines for applications including Rooftop and Obstacle Detection, Segmentation, Face Recognition, and GANs across various computing platforms. I am well-versed in Deep Learning libraries, and GenAI/ML toolkits such as PyTorch, TensorFlow, Keras, MCP, A2A, OpenCV, and PIL, and proficient in programming with Python, C/C++, SQL, R, PHP Laravel, MATLAB, and ROS. My project portfolio, including Neural Machine Translation and Domain Adaptation, showcases my ability to tackle complex AI challenges.
Outside the tech world, I am passionate about photography and music production, which enhance my creative outlook. As a friendly and enthusiastic professional, I welcome discussions and collaborations on topics like machine learning, computer vision, deep learning, and advanced algorithms. If you share these interests, I encourage you to reach out so we can explore innovative solutions together.
• Led Nexus, a multi-agent GenAI compliance platform for 37 regions and 190 countries; enabled natural-language rule discovery and full CRUD and reduced discovery/authoring time by 90% (from about 2 weeks to 2–3 minutes, typically 1–2 queries).
• Architected dual access paths—semantic RAG web and a Claude-powered CLI (MCP client + custom servers); bash-style commands enable zero-vector, Git-backed traversal and custom creation of stateless agents for subtasks; unified 4 sources (registries, Git JSON, legacy Java rule logic).
• Built scalable and audit-ready pipelines using LangGraph agents for classification, retrieval, and validation, plus versioning and audit trails; technologies: Python, FastAPI, LangGraph, Claude, FastMCP, Docker, GitHub Actions.
• Engineered an automated annotation pipeline using MediaPipe’s PoseLandmarker to extract and visualize infant posture keypoints, reducing manual labeling time by over 50%
• Curated 100K+ infant posture images from baby monitor videos, manually annotating 0.37% while leveraging a custom Python script for bulk auto-annotation, streamlining dataset creation
• Improved infant pose estimation accuracy by refining data ingestion workflows, integrating segmentation masks, and iterating on an invariant representation learning model (IEEE FG, 2021) for robust infant posture tracking
• Boosted solar project estimation efficiency by 50% by developing and productionizing machine learning algorithms using Python and Django. Processed location data through a Glue ETL pipeline into a MySQL database, leveraging Hadoop-Spark architecture to significantly reduce errors
• Increased decision-making efficiency by 40% by architecting and integrating comprehensive financial analytics APIs in Laravel. Streamlined calculations for ROI and payment scenarios, enhancing the platform's analytical capabilities
• Increased roof detection accuracy by 30%, processing speed by 40%, and client engagement by 35% by integrating ResNet, optimizing data pipelines, and developing imaging solutions and optimal panel placement algorithms using OpenCV and managing data with AWS S3
• Performed sentiment analysis on 15,000 British Airways customer reviews (BeautifulSoup, Pandas, NLTK, VADER), revealing that about 20% of comments highlighted in-flight amenities, enabling targeted improvements aligned with traveler preferences
• Developed and optimized Random Forest and XGBoost models to predict rebooking tendencies, elevating accuracy from around 60% to nearly 78%, supporting data-driven allocation of routes and resources
• Designed interactive dashboards (Seaborn, Power BI, Tableau) to visualize sentiment shifts and recurrent themes, reducing manual review time by approximately 30% and accelerating stakeholders’ response to emerging feedback patterns
• Supervised a team of six to develop a road obstruction detection model using CNN MobileNet, achieving 80% accuracy on dashboard footage and enhancing road safety for autopilot applications
• Increased data annotation efficiency by 50% and reduced manual annotation time by 30% by streamlining processes with AuVi.io, Edgematrix.io, and Makesense.ai, accelerating model development cycles
• Improved model runtime by 67% and boosted accuracy to 97% on edge devices by implementing TensorFlow and TensorFlow Lite for object detection. Reduced model size by 40% (from 240MB to 80MB) through pruning and quantization of the YOLOv5 model, enhancing real-time application performance
Related courses: Algorithms, Computer Vision, Advanced Perception, Machine Learning, Robotic Science and Systems, Foundations of Artificial Intelligence, Advances in Deep Learning, Database Management Systems
Related courses: Image Processing, Web Programming, Data Structures, OOP, Computational Mathematics
See My Work - Projects
This project employs domain adaptation techniques to enhance a deep learning model trained on the SVHN dataset for digit recognition. By adapting the model to the MNIST dataset, it improves performance on the target domain with limited labeled data.
View Project
Detecting motion using Simple Image Filtering between sequential video frames using Gaussian, 3x3 Box, and 5x5 Box filters.
View Project
Focused on neural machine translation (NMT), this project utilizes Bahdanau attention, encoder-decoder architecture, and deep learning. Trained on Stanford datasets for English to Vietnamese and English to Dutch, the NMT system generates accurate translations preserving context and structure.
View Project
This project employs SIFT for keypoint detection and matching between two images using the FLANN matcher and Lowe's ratio test. It provides visual outputs of keypoints, transformations, and depth perception through a disparity map.
View Project
A learning project to implement a basic voice assistant using Python, nothing complicated.
View Project
This C++ script implements a task scheduling algorithm for Mobile Cloud Computing (MCC), optimizing energy usage and meeting deadlines by dynamically allocating tasks between local devices and the cloud. Based on research by Xue Lin et al., it enhances performance in MCC environments through task migration and dynamic prioritization.
View Project
This Personal Assistant Chatbot project leverages Streamlit, OpenAI's GPT-3.5, and various APIs to provide an intelligent conversational assistant. Key features include document processing, web search, weather updates, and the implementation of Retrieval-Augmented Generation (RAG) for enhanced response accuracy and relevance.
View Project
This project uses reinforcement learning to train an agent to play Flappy Bird using the Deep Q-Network (DQN) algorithm with the Stable Baselines3 library. The environment is built with OpenAI Gym and Pygame.
View Project
G.R.O.O.T is an end-to-end pipeline for 30-day hospital readmission prediction using MIMIC data. It combines structured features with on-the-fly text embeddings in a Fusion Model or XGBoost, and provides a Retrieval-Augmented Generation (RAG) module to produce short medical summaries for enhanced interpretability.
View Project
San Jose, CA, USA
+1 (617) 817-6535
dhanushbalakrishna45@gmail.com