projects | Raghu Hemadri

work

Real-Time Healthcare Chatbot with End-to-End MLOps (Dr. Dialog)

Dr. Dialog is a fault-tolerant, MLOps-enabled healthcare assistant built on TinyLLaMA-1.1B, designed to simplify medical jargon and answer non-personalized health queries in real time. The system features distributed GPU training, continuous delivery with ArgoCD, and robust monitoring via Grafana and Prometheus.

Parameter-Efficient Deep ResNet with Channel Attention

Lightweight ResNet variant with Squeeze-and-Excitation attention for CIFAR-10, achieving 96.04% test accuracy under a strict 5M parameter budget. Integrated Mixup, Cutout, AutoAugment, and cosine-annealed SGD for state-of-the-art regularization and convergence.

Parameter-Efficient Transformer for Scalable NLP Classification

Parameter-efficient fine-tuning method using Low-Rank Adaptation (LoRA) on RoBERTa for news text classification. Achieved 86.5% test accuracy with under 1M trainable parameters, demonstrating near state-of-the-art performance at a fraction of computational cost.

Adversarial Robustness of Deep Image Classifiers (ResNet-34, DenseNet-121)

Investigated the vulnerability of ResNet-34 and DenseNet-121 classifiers to adversarial attacks using FGSM, PGD, and a custom patch-based PGD method. Demonstrated near-total degradation of classification accuracy with transferable adversarial examples and provided detailed ablation studies.

Reasoning with LLMs for Cancer Treatment Outcome Prediction

A reasoning-aligned LLM framework for predicting cancer treatment outcomes using the MSK-CHORD dataset, integrating Chain-of-Thought prompting and reinforcement learning. Achieved state-of-the-art performance in both accuracy and interpretability for clinical decision support.

Age and gender estimation

A CNN-based system using TensorFlow and dlib for accurate age and gender estimation from facial images, enabling multi-face demographic analysis.

Rally Master AI

An AI-driven badminton training analysis system using perspective transformation and bird’s-eye view analysis to provide data-driven posture correction and performance assessment.

Log File Summarization

A log file summarization tool using the Drain algorithm to extract key insights and patterns, improving system monitoring and troubleshooting efficiency.

3D Scene Reconstruction

Accurate 3D maps using for multi-view scene reconstruction.

FitGen - Personalized Exercise Planner

Generate personalized exercise plans, optimizing workouts through real-time form feedback to enhance fitness efficiency and reduce injury risk.

PyTorch-Powered Machine Learning Models

A comprehensive repository of machine learning models in PyTorch, providing well-structured implementations for applications in computer vision, NLP, and beyond.

fun

Social Neuro RL

Explores context-conditioned reasoning and compositional thinking in AI agents using MARL to enhance coordination, generalization, and cognitive architectures inspired by social neuroscience.

MAGNET: Multi-Agent Generative Network

Multi-Agent Generative Network, a dual-agent cooperative architecture for data generation, outperforming traditional GANs in producing high-quality images resembling the CelebA dataset.

Meta Learning with exact information estimation

An information-theoretic meta-regularizer to mitigate overfitting in meta-learning by maximizing mutual information using MINE, improving generalization on non-mutually-exclusive tasks.

Covivac Bot

A telegram bot for real-time vaccine slot availability checks and instant alerts, enabling seamless and efficient appointment booking.

Reinforcement Learning with Unity

A RL project using Unity ML Agents, training intelligent agents in virtual environments for applications in gaming, robotics, and simulations.

Stock Trading using RL

A RL-based stock trading system, training RL agents on historical data to optimize portfolio management and enhance market returns.