📝 Publications

HalluMat: Detecting Hallucinations in Domain-Specific Large Language Models for Materials Science
Bhanu Prakash Vangala, Jianlin Cheng

• Key Innovation: Introduced HalluMatData benchmark dataset and HalluMatDetector framework for comprehensive hallucination detection in scientific LLMs
• Technical Contributions: Multi-stage pipeline combining intrinsic verification, multi-source retrieval, contradiction graph analysis, and metric-based assessment
• Impact: Achieved 30% reduction in hallucination rates compared to baseline LLM outputs in materials science domain
• Novel Metric: Introduced Paraphrased Hallucination Consistency Score (PHCS) to quantify response inconsistencies across semantically equivalent queries

HalluFormer: Transformer-based Faithfulness Evaluation Framework for Large Language Models
Bhanu Prakash Vangala, Jianlin Cheng

• Architecture: Transformer-based model for multi-dimensional consistency checking between questions, answers, and retrieved knowledge contexts
• Performance: Achieved F1 score of 0.9471 on MultiNLI test dataset and 0.7285 on blind ANAH test dataset
• Generalization: Demonstrates strong cross-domain generalization capabilities for detecting LLM hallucinations
• Applications: Deployed in critical scientific and clinical domains requiring high factual accuracy

Deploying LLM-as-a-Service in Kubernetes HPC Clusters
Bhanu Prakash Vangala, Grant Scott, Jianlin Cheng

• Infrastructure Innovation: Designed Helm-based GPU-aware deployment pipeline for seamless LLM inference in research clusters
• Scalability Features: Implemented auto-scaling, resource throttling, and multi-tenant access configurations
• Performance: Reduced deployment time by 60% and improved resource utilization by 40% in HPC environments
• Real-world Deployment: Successfully deployed on Nautilus cluster serving 100+ concurrent users

Brain Tumor Detection in MRI Images Using Deep Learning
Bhanu Prakash Vangala

• Model Architecture: Custom CNN and ResNet50V2 with advanced data augmentation techniques
• Performance Metrics: Achieved 96.8% accuracy, 0.95 F1-score on brain tumor classification
• Technical Innovations: Implemented learning rate optimization and ensemble methods for improved robustness
• Clinical Impact: Developed interpretable visualization techniques for medical practitioners

Pneumonia Detection in Chest X-rays Using Deep Learning
Bhanu Prakash Vangala

Project Report

• Model Comparison: Evaluated 5 architectures including custom CNN, ResNet18, VGG16, ResNet50 with K-Fold CV, and EfficientNet
• Best Performance: EfficientNet achieved highest classification metrics with 94.2% accuracy
• Training Innovations: K-fold cross-validation and multi-GPU acceleration for robust model training
• Dataset: Trained on 5,216 chest X-ray images with rigorous validation protocols

AI-Powered Image Colorization: From Grayscale to Photorealistic Color
Bhanu Prakash Vangala

Paper

• Architecture: Convolutional autoencoders and GAN-based models for photorealistic colorization
• Innovation: Novel loss functions combining perceptual and adversarial losses for natural color generation
• Applications: Successfully applied to historical photograph restoration and film remastering
• Performance: Achieved SSIM score of 0.92 and user preference rating of 87% in blind studies

Multilingual Sentiment Analysis on KOO: Uncovering User Sentiments and Trends
Bhanu Prakash Vangala

• Multilingual Support: Developed sentiment analysis pipeline supporting 10+ Indian languages
• Real-time Processing: Implemented streaming architecture for real-time sentiment detection
• Scale: Processed over 1M posts with sub-second latency for content moderation
• Impact: Deployed in production for enhanced user experience and safety on social platform