Vara Prasad Gudi
Computer Scientist
Computer Scientist
A goal-oriented computer scientist who loves turning complex challenges into clear, effective solutions by blending machine learning, data science, and software engineering. I thrive in collaborative environments where innovative ideas drive practical results, and I enjoy creating systems that simplify processes and make technology accessible. My experience spans research and hands-on industry work, and I'm always exploring new ways to harness tech for meaningful change. I'm excited to connect with forward-thinking teams and leaders who are ready to transform ideas into impact.
I finetuned LLaMA and DeepSeek-R1 with QLoRA to reduce hallucinations in medical summaries. I also developed a dual-agent LLM inference pipeline using Python, Langchain, and Hugging Face models to enhance both detection and mitigation processes. Currently, I'm implementing GRPO and self-refinement strategies to design an automated metric for mitigating medical hallucinations.
I designed an LLM agent using LangChain, GPT-4, and Map-Reduce workflows for survey theme extraction, employing semantic chunking and summarization techniques to drastically streamline analysis. I also created a robust FastAPI microservice in Python, integrating custom-trained YOLOv11 and EasyOCR models on Databricks to automate distribution and packaging workflows. Additionally, I developed a full-stack batch record management system using Next.js, TypeScript, and PostgreSQL, incorporating real-time Kanban tracking, Power BI dashboards, and auto-scaling features to efficiently manage records. Furthermore, I engineered ETL pipelines with PySpark, Databricks, and SQL for big data processing in TrackWise Oracle DB, enabling the delivery of dynamic Power BI dashboards for daily tracking.
I spearheaded the design and evaluation of robust NLP models for out-of-distribution sentiment analysis across multiple datasets by implementing innovative techniques such as K-shot learning, chain-of-thought prompting, and in-context learning. I also curated benchmarks in medical NLP using Gaussian Mixture Models and UMAP to identify distribution shifts. Additionally, I constructed end-to-end LLM pipelines using state-of-the-art models like BERT, Llama, and the GPT family, processing large-scale training samples and optimizing GPU memory usage through advanced quantization techniques, resulting in strong in-distribution performance.
I focused on optimizing delivery and supply chain processes by programming a backtracking algorithm in Python to improve cruise schedules. I trained machine learning models with Azure ML and Python for supply-demand forecasting and designed ETL workflows for SAP data extraction, which supported comprehensive Power BI reporting. I also analyzed regional demand patterns by joining sales, inventory, and weather data, enabling faster responses to operational changes.
Developed and containerized microservices using PHP (Laravel), Docker, and Kubernetes for a scalable video management system with YouTube API integration. Implemented real-time video communication leveraging React, WebRTC, and Agora API to enhance platform responsiveness and user experience.
Published a chapter in the book "Explainable Artificial Intelligence (XAI): Concepts, Enabling Tools, Technologies and Applications," published by The IET Digital Library, provides a comprehensive insight into the role of XAI in the medical field. With its Chapter DOI: 10.1049/PBPC062E_ch, it emphasizes the necessity for transparency in AI systems among healthcare professionals. The chapter thoroughly explores applications in areas such as medical image analysis with deep learning, clinical decision support, and broader healthcare. It also highlights XAI's role in enhancing trust in healthcare, with in-depth discussions on its integration in healthcare frameworks for pandemic prevention, and specific medical applications like allergy diagnosis. This research significantly contributes to the understanding of how XAI can be a transformative force in medical science, emphasizing the importance of explainability in the deployment of AI solutions for improved health outcomes.
This research, conducted under the guidance of Assistant Prof. Sheela Jayachandran at the School of Computer Science and Engineering (SCOPE) at VIT-AP, presents a groundbreaking Python-based machine learning algorithm for image analysis. The Fuzzy Image Clustering via Featured Interval Extraction algorithm excels with a 100% accuracy rate in overlap distance evaluation, significantly outperforming existing methods in datasets with high overlap. Its versatility and precision in multivariate statistical applications not only mark a significant advancement in the field but also earned recognition through a utility patent (IP India Application No. 202341029896). This work highlights the potential for enhancing image analysis techniques and contributes substantially to the research community.
This research, under Assistant Prof. Sheela Jayachandran at VIT-AP's School of Computer Science and Engineering, has significantly advanced waste management technology. By integrating the YOLOv4 algorithm with IoT, the team developed a system that not only surpasses previous models in accuracy but also improves efficiency. This innovation led to a notable 10% increase in detection precision and a 12% boost in processing speed, showcasing a scalable solution that melds deep learning with environmental sustainability. The impact of this groundbreaking work is further acknowledged through the publication of a design patent (IP India Application No. 202341070187), marking a milestone in sustainable technology development.
This research, led by Associate Prof. Somya Ranjan Sahoo at VIT-AP's SCOPE, innovatively applies fuzzy logic to traffic control systems, resulting in a 30% improvement in traffic flow and a 20% reduction in vehicle wait times. The study involved a Pygame simulation of a complex four-way intersection, showcasing its potential in urban traffic management. The team's findings, demonstrating a 45% efficiency increase over static systems, were recognized for their significance in urban planning and smart cities, culminating in a publication in an IEEE journal.