DHRUV.
KUMAR

Indian Institute of Technology Kharagpur

Kharagpur, West Bengal, India

Jul 2025 – Present Contract

Junior Project Engineer

• Authored an ISRO research proposal on interpretable crop health monitoring and early disease detection using hyperspectral satellite imagery, proposing a contrastive learning framework to address satellite–ground resolution mismatch.
• Implemented RGB-to-Hyperspectral reconstruction architectures to reduce hyperspectral acquisition costs and leveraged Google Earth Engine for large-scale agricultural data extraction and preprocessing.
• Developed a multimodal persuasive image ranking framework by fine-tuning vision–language models on a 450-group human-evaluated benchmark dataset, resulting in a research manuscript currently in preparation.
• Implemented and benchmarked 7 state-of-the-art Hindi machine-generated text detection models on curated and paraphrased datasets generated using Gemini, Llama, Qwen, and GPT-4o, enabling cross-LLM evaluation across multiple generation scenarios; results documented in a manuscript under review at SIGIR 2026 Reproducibility Track.
• Designed a hybrid Hindi MGT detection architecture combining middle-layer embeddings with KL-divergence signals from two auxiliary LLMs, achieving ~95% detection accuracy across evaluation setups.

Prodigal AI Technologies

New Delhi, India

Sep 2024 – Dec 2024 Internship

Machine Learning Intern [PyPI] [Source Code] [Web App]

• Accelerated LinkedIn content production by 40% by developing Co-Agent, an open-source multi-agent conversational framework using Google Gemini API for blog-to-post transformation.
• Processed 6,000+ blog posts via full-stack NLP pipeline: Web Scraping, Data Cleaning, API Integration, Summarisation, AI Agent-based Validation, and LinkedIn-specific Formatting.
• Managed a 5-intern team for R&D; launched a PyPI package with 3,500+ downloads in 3 months, backed by a proof-of-concept WebApp.

ISRO – SDSC SHAR

Sriharikota, Andhra Pradesh, India

Jul 2024 – Aug 2024 Internship

Machine Learning Intern [View Website]

• Achieved 97.8% accuracy (↑10%) and 95.3% recall by designing a "Multi-model ML Architecture" ensembling ANN & CNN for space-domain classification.
• Reduced false negatives to 4.7% via class balancing, EDA, feature engineering, normalisation, and augmenting 90K+ samples.
• Deployed a scalable Research-Aid SaaS with real-time API integration via Django REST Framework in collaboration with 5 Astronomers, a full MLOps pipeline.