Vikas Chelluru

• Engineered a state-of-the-art multi-agent framework featuring an advanced SQL agent that interfaces with BigQuery and Vertex AI Search through a RAG pipeline for highly accurate Results. Deployed the system on AWS ECS with a FastAPI backend for robust, scalable performance. • Developed and fine-tuned a specialized code-generation agent for the ’Verse’ gaming language, leveraging RAG to enhance code understanding and accelerate game development. The agent is integrated into a FastAPI backend with WebSockets for a seamless, interactive user experience. • Architected a sophisticated research agent for trend analysis, utilizing Tavily for broad searches and Bright Data for deep, targeted data extraction. The agent autonomously generates structured JSON reports, providing actionable insights from complex information landscapes. • Developed Gen AI pipelines, including text-to-image and image-to-image workflows with fine-tuned LoRAs for consistent custom character and style generation. Also engineered a high-fidelity voice cloning pipeline and an automated Instagram Reel generator, deploying these solutions on RunPod for serverless, on-demand processing.

In my role as an AI Engineer, I specialize in developing cutting-edge AI solutions in the fields of computer vision and generative AI. My key responsibilities and achievements include: AI-Driven Patient Monitoring Systems: Designed and implemented advanced patient monitoring systems using MediaPipe Pose, YOLOv8, and YOLOv8X-Pose for activity recognition and safety monitoring. • Architected and deployed a high-throughput, real-time patient monitoring system, processing concurrent CCTV streams via a robust FastAPI-based pipeline. Engineered a microservices architecture on AWS ECS (Fargate) and Lambda to deliver critical services including fall detection, out-of-bed alerts, and automated scene descriptions. • Developed an agentic AI chatbot using LangGraph and AWS Bedrock to provide contextual insights on patient activity. The agent efficiently queries historical stream logs and alert data from DynamoDB, featuring effective memory management for coherent, long-term conversations with users. • Enhanced medical posture classification accuracy by 38 percentage by applying parameter-efficient fine-tuning (PEFT) techniques (LoRA/QLoRA) to vision-language models like Qwen2.5-3B and OpenCLIP, enabling highly efficient model adaptation. • Spearheaded the development of a multi-modal Medical Assistant Agent for analyzing diverse medical images (e.g., X-rays). The agent leverages AWS Knowledge Bases to create a robust Retrieval-Augmented Generation (RAG) architecture, integrating multiple CV models and LLMs with critical safety safeguards. • Engineered the core cloud infrastructure for GenAI and ML operations, utilizing AWS ECS for container orchestration, AWS SageMaker for optimized inference, and AWS Bedrock for foundation model access. Integrated DynamoDB for scalable data logging and AWS SNS to trigger real-time alerts, ensuring proactive patient safety and system reliability

During my tenure as a Computer Vision Intern, I was deeply involved in several advanced computer vision projects and played a key role in developing an end-to-end application. My responsibilities and achievements include: End-to-End Application Development: Developed an application that processes live CCTV camera streams for real-time object detection and tracking, tailored for the manufacturing industry to enhance resource management, automatic record-keeping, safety monitoring, and violation detection. Algorithm Optimization and Development: Improved the efficiency and accuracy of object detection and tracking models, ensuring low latency and high performance. Backend Development with Django: Utilized Django for backend development, ensuring robust server-side operations, and integrated PostgreSQL for efficient storage and retrieval of real-time information. Model Training and Architecture Modification: Engaged in model training and architecture modification to fine-tune performance, experimenting with different neural network architectures. Live Stream Processing: Focused on processing live streams, employing multithreading techniques to handle processes separately for smooth operation. Modular Coding: Adopted a modular coding approach to enhance code maintainability, scalability, and collaboration. Frontend Dashboard Creation: Developed a user-friendly dashboard displaying live inference videos, statistics through graphs and tables, and direct database access for users. Cross-Disciplinary Collaboration: Collaborated with software engineers, data scientists, and project managers to align technical solutions with business needs, participating in brainstorming and problem-solving sessions. Learning and Adaptation: Continuously learned and applied new computer vision techniques, staying updated with the latest advancements in the field. Note: Designed and implemented a real-time data pipeline ensuring smooth live stream processing and data handling.

I Worked to develop the solutions for the yield detection in mango farms by capturing the 360 degree images using UAVs. The proposed approach will utilize technological advancements such as UAVs, Internet of Things (IoTs), Computer Vision (CV), etc. Primary objectives of the project are to collect the data from the mango farms using the UAVs equipped with multiple sensors. The collected data will be processed with the help of object detection techniques for the mango fruit detection which is required to measure the yield in mango farms. My work : Intel RealSense Camera on Raspberry Pi for Mango Depth Analysis: Implemented Intel RealSense depth camera on Raspberry Pi for mango depth analysis. Conducted data collection in mango fields for precise yield estimation. Utilized Raspberry Pi mounted on a drone for real-world applications. Skills Acquired: Developed expertise in Deep Learning, machine learning, Computer Vision, IoT, and MATLAB/Simulink. Applied acquired skills to create innovative solutions for complex challenges. WiFi Communication Performance Experiments: Conducted experiments on WiFi communication between NodeMCUs and Raspberry Pis. Explored data rate variations in different environments and ad-hoc networking scenarios. Helipad Detection and Automatic Charging Project: Worked on helipad detection and automatic charging of drones using computer vision. Collected and processed drone camera data for model training. Utilized deep learning models like YOLO and SSDMobileNet v2 FPNLite 640 for detection. Thermal Imaging for Mango Fruit Condition Classification: Used thermal camera technology for collecting data on mango fruit conditions. Conducted preprocessing on thermal profile images and applied machine learning and deep learning techniques. Ongoing Paper Work: Currently involved in writing a paper based on the helipad detection and automatic charging project. Demonstrating in-depth understanding of computer vision and deep learning concepts.

Dedicated myself to mastering advancements in the communication field during my internship. Explored complex concepts by studying the latest research papers with a focus on computational mathematics. Guided by Professor Divyabhramam, leveraged his thesis to derive equations and gain unique insights. Engaged collaboratively with Ph.D. students, visualizing complex concepts and receiving appreciation for my work. Proficiently used MATLAB for result validation, data analysis, and creating visual representations. Overcame challenges in the cutting-edge project with proactive problem-solving and mentor guidance. Delivered innovative solutions to complex problems, showcasing resourcefulness and critical thinking. Accomplished two significant tasks within 2.5 months, contributing to the project's overall objectives. Eagerly seeking new opportunities to apply gained skills and knowledge, demonstrating a passion for learning and innovation. Developed strong problem-solving skills through hands-on experiences during the internship.