class VedantPandhare:
def __init__(self):
self.name = "Vedant Pandhare"
self.role = "AI Engineer"
self.location = "India 🇮🇳"
self.languages = ["Python", "C++", "JavaScript", "R"]
self.interests = ["Computer Vision", "NLP", "LLMs", "RAG"]
self.currently = "Building AI-driven real-world apps"
def say_hi(self):
print("Thanks for visiting! Let's build something amazing.")
me = VedantPandhare()
me.say_hi()LLM Compression · Multilingual NLP · Transformer Pruning · Model Efficiency · Block Influence
arXiv Preprint — 2026
- Investigates how the language used for calibration affects Block Influence (BI)-based depth pruning in multilingual language models.
- Evaluated Qwen2.5-1.5B-Instruct, Llama-3.2-1B, and Gemma-2-2B across English, Hindi, Marathi, and Tamil.
- Found that late Transformer blocks show 20.7–55.5% higher Block Influence for Hindi, Marathi, and Tamil than for English, showing that English-only calibration can underestimate the importance of late layers.
- Despite high English-to-target ranking correlations, the actual sets of blocks selected for pruning can still differ substantially.
- Evaluated equal-budget language-balanced calibration over English, Hindi, and Marathi, while keeping Tamil held out to test cross-script transfer.
- Results show improvements for Qwen and Gemma in several pruning settings, while Llama provides an important counterexample, showing that multilingual balancing is not universally robust.
- The work highlights that calibration language should be treated as part of the pruning algorithm itself, rather than as an incidental implementation choice.
🔗 arXiv: 2610.21383
Deep Learning · CNN · YOLOv8 · OpenCV · Medical Imaging
- Developed a deep learning pipeline to detect and classify intervertebral disc degeneration across multiple disease categories.
- The study explores the use of Convolutional Neural Networks (CNNs) to identify radiographic patterns associated with disc degeneration, including disc-space narrowing and osteophyte formation.
- Trained and evaluated models on 10,075 annotated spinal X-ray images for automated medical image analysis.
- By leveraging automated medical image analysis, the work investigates how deep learning can assist in detecting degenerative changes and provide a more efficient and consistent approach to radiographic assessment.
- Peer-reviewed research presented at ICAISI-25 International Conference.
🔗 DOI: 10.1201/9781003654049-50
NLP · K-Means Clustering · NER · Hugging Face
- This research presents a novel AI framework for automated PDF summarization that addresses the challenge of efficiently processing large volumes of textual information.
- The study explores K-Means clustering approach, Sentence Transformers, with new extractive summarization, and abstractive summarization methods to reduce document length by 70% while retaining essential semantic information.
- The work demonstrates the potential of machine learning and NLP techniques for scalable document analysis and efficient knowledge extraction.
- Published in the Proceedings of ICSIAIML-25, Atlantis Press.
🔗 DOI: 10.2991/978-94-6463-948-3_50
AI/ML Intern
- Developed a GraphRAG pipeline modeling 3+ entity-relationship layers across political networks, fusing a knowledge graph with vector-based semantic retrieval to surface non-obvious connections for sentiment and policy analysis.
- Engineered an LLM-powered evaluation & benchmarking system for AI calling workflows, defining 5+ validation metrics and regression tests to extract citizen sentiment toward elected officials and generate actionable policy insights.
- Deployed a RAG-based internal chatbot over a real-time news vector database, applying chunking and embedding strategies across 2+ retrieval pipelines with Langfuse observability, enabling leadership to access context-aware briefings 3× faster.
Machine Learning Intern
- Architected a real-time anomaly detection and fault prognostics system using multi-axis IoT sensors and ensemble ML models; isolated failure modes like mechanical looseness and rotor dynamics imbalances, slashing production halts and false alarms by 65%.
- Implemented advanced time-series forecasting utilizing LSTMs, ARIMA, and Prophet to model asset degradation curves and predict Remaining Useful Life estimation (RUL) long before physical symptoms occurred.
- Engineered an edge-AI data acquisition pipeline featuring digital signal processing (FFT) to extract high-frequency vibration signatures, enabling scalable, real-time predictive maintenance on industrial equipment.
AI Student Intern
- Developed a secure AI inference pipeline for a defense-grade drone system
- Integrated Raspberry Pi with Hailo 8L AI accelerator for efficient edge deployment
- Bridged hardware and AI capabilities for mission-critical applications
