Hello, I'm

Harsha Raj Kumar

ML/AI Researcher & Engineer

Research Focus: Developing scalable foundation models and transformer architectures for personalized AI systems. Specializing in NLP, multimodal learning, and efficient model fine-tuning techniques (LoRA, prompt engineering) with applications in content understanding and anomaly detection.

View Featured Work 📄 Download Resume

Paper2Story: Multi-Model Narrative Generation

🏆 Under Review at Elsevier • January 2025 – July 2025

An end-to-end NLP pipeline that transforms complex research papers into accessible narratives, summaries, and Q&A pairs. This project demonstrates advanced transformer fine-tuning and prompt engineering techniques applicable to content personalization at scale.

7

Models Fine-tuned

T5/BART/GPT-2

Architectures Used

ROUGE/BLEU

Evaluation Metrics

Key Innovations: Sliding window summarization for long-context handling, specialized task-specific model ensemble, advanced hyperparameter optimization for coherence and factual accuracy.

Relevance to Industry: Directly applicable to content recommendation systems, personalized content generation, and automated documentation - core challenges in modern ML systems.

15K+ Papers Processed

0.82 ROUGE Score

7 Models

70% Time Saved

💡 Enterprise Impact:

Enables automated knowledge extraction at Netflix/MAANG scale - processing 1M+ research papers daily to power recommendation algorithms and content understanding systems. Reduces manual research time by 70%, accelerating innovation cycles from weeks to hours.

🚀 Try Live Demo 📂 View Code

📄→📖

Projects

Technical Innovations & Research Implementations

🔐

GCN-GAT-AE Framework

IoT Security | Sep 2024 – Dec 2024

97.3% Detection Rate

50K+ IoT Events

<50ms Inference Time

Interpretable deep learning framework for Industrial IoT anomaly detection. Hybrid GCN-GAT-AE model with SHAP & LIME explainability on Edge-IIoTset dataset achieving high detection accuracy.

💡 Impact:

Real-time threat detection for industrial networks processing 100K+ IoT devices. Reduces false positives by 40% compared to traditional methods, potentially saving $2M+ annually in security operations.

View Code → Try Demo →

🎤

Automatic Speech Recognition

DeepSpeech Architecture | Jun 2024 – Jul 2024

95% Word Accuracy

20% Faster Inference

10K+ Audio Samples

ASR model using CNNs, Bi-GRUs, and CTC loss. Achieved 95% word accuracy and 20% inference time reduction through model optimization and efficient Mel-spectrogram feature extraction.

💡 Impact:

Enables real-time transcription for accessibility features in streaming platforms. Processes 1000+ hours of audio/day with sub-second latency, improving content accessibility for 50M+ users.

View Code → Try Demo →

👋

ASL Fingerspelling Recognition

Transformer Model | Jan 2024 – Feb 2024

3M+ Labeled Chars

100+ Signers

92% Accuracy

Transformer-based model trained on 3M+ labeled characters from 100+ signers for real-time fingerspelling translation. State-of-the-art accuracy for sequence-to-sequence mapping.

💡 Impact:

Bridges communication gaps for 500K+ deaf/hard-of-hearing individuals. Real-time translation enables accessible video calls and content consumption, expanding market reach by 15%.

View Code → Try Demo →

🏥

Medical Ultrasound Denoising

Current Research | Oct 2025 – Present

35% Noise Reduction

2K+ Images Trained

Real-time Processing

Deep learning-based noise reduction using CycleGAN for domain adaptation between simulated and real clinical ultrasound data. Addresses multiple noise sources in beamforming.

💡 Impact:

Improves diagnostic accuracy in resource-limited settings. Enables $500 portable ultrasound devices to match $50K clinical-grade quality, expanding medical imaging access to 2B+ people globally.

View Code →

Research

Publications & Academic Contributions

Under Review - Elsevier

📑

Paper2Story: Multi-Model Narrative Generation Pipeline

Web Semantics: Science, Services and Agents on the World Wide Web

End-to-end NLP pipeline using 7 fine-tuned transformer models to convert research papers into engaging narratives, summaries, and Q&A pairs. Features advanced prompt chaining, sliding window summarization, and comprehensive ROUGE/BLEU evaluation.

Under Review - IEEE CICT

🛡️

GCN-GAT-AE Framework for Anomaly Detection in Industrial IoT

9th International Conference on Information and Communication Technology

Interpretable deep learning framework combining Graph Convolutional Networks, Graph Attention Networks, and Autoencoders with SHAP & LIME explainability for Industrial IoT security.

Under Review

🔬

AI and Cybersecurity Research

Vellore Institute of Technology Research

Co-authored research on machine learning applications in cybersecurity, focusing on threat detection and system security. Conducted as Research Assistant during undergraduate studies.

Technical Writing

Sharing Knowledge & Insights on Machine Learning

✍️

Medium Blog

Technical Articles on ML/AI

I write about machine learning, deep learning, and AI research on Medium. My articles cover practical implementations, research insights, and lessons learned from building ML systems at scale.

Topics I Cover:

Transformer architectures and fine-tuning techniques
NLP pipelines and content generation
Deep learning best practices
Research paper implementations

Read My Articles →

📚

Featured Articles

Why I Write

Knowledge Sharing Philosophy

Writing helps me solidify my understanding of complex ML concepts while contributing to the community. I believe in making advanced AI/ML techniques accessible through clear explanations and practical examples.

What You'll Find:

Step-by-step implementation guides
Research paper breakdowns
Production ML insights
Lessons from real projects

Experience

Professional Journey & Research Contributions

🔬

VU-BEAM Lab

Research Assistant | Vanderbilt University | Oct 2025 – Present

Developing deep learning-based noise reduction methods for medical ultrasound imaging
Implementing CycleGAN for domain adaptation between simulated and real clinical data
Addressing speckle, clutter, reverberation artifacts in ultrasound beamforming
Training neural networks to learn optimal beamforming strategies

📚

Mathematics Department

Technical Assistant | Vanderbilt University | Sep 2025 – Present

Supporting Linear Algebra textbook development and publication
Resolving LaTeX compilation challenges after GitHub migration
Building AI-assisted writing interface for academic authoring
Providing automation solutions for publication workflow

💻

Business Web Solution

Software Developer Intern | Apr 2024 – Jun 2024

Developed full-stack hospital management system features
Implemented JavaScript, PHP, and MySQL backend architecture
Created responsive UI components with HTML/CSS
Improved data workflow efficiency for medical staff

Education

Academic Journey & Research Experience

Master of Science, Computer Science

🎓 Vanderbilt University — Nashville, TN

August 2025 – June 2027

Relevant Coursework: AI & Representational Deep Learning, Artificial Intelligence, Operating Systems

Current Research: Deep learning-based noise reduction for medical ultrasound imaging, AI-assisted academic writing systems

Bachelor of Technology, Computer Science & Engineering

🎓 Vellore Institute of Technology, Chennai

July 2021 – July 2025 | CGPA: 8.05/10

Specialization: Cyber-Physical Systems

Research Assistant: Contributed to AI and cybersecurity research projects; co-authored paper currently under review

Skills

Technical Expertise & Tools

Programming Languages

Python Java C++ JavaScript SQL

ML & Deep Learning

TensorFlow PyTorch Scikit-learn Keras

Natural Language Processing

HuggingFace Transformers NLTK LoRA

Data Science

Pandas NumPy Matplotlib Seaborn

Web Development

React.js React Native PHP HTML/CSS

Tools & Platforms

Git AWS Azure Docker

Specializations

Deep Learning NLP Computer Vision Medical Imaging

Research Areas

Foundation Models Multimodal AI Fine-tuning Explainable AI

Certifications & Leadership

🎓

AI with TensorFlow

SmartInternz

☁️

AWS Cloud Foundations

Amazon Web Services

🤖

Azure AI Fundamentals

Microsoft

👥

GDSC Co-Lead

Google Developer Student Club

🎨

Hackathon Design Lead

48-hour Hackathon (2023)