Atul Deshpande

Recent graduate focused on low-resource natural language processing, information retrieval, and applied artificial intelligence in healthcare. Committed to reproducible research and open science.

View Research Get In Touch

About

I am a recent graduate with experience spanning natural language processing, information retrieval, and computational mechanics. My research focuses on developing efficient, reproducible methods for complex systems, ranging from low-resource language modeling to diagnostic applications in healthcare.

I am currently preparing for Master's degree programs, seeking opportunities to contribute to research in intelligent systems and applied AI.

Research Interests

Natural Language Processing

Developing robust language models and evaluation frameworks for low-resource and multilingual contexts, focusing on data-efficient adaptation.

Information Retrieval

Investigating cross-lingual search systems, open benchmark creation, and semantic retrieval architectures using vector databases.

Reasoning in LLMs

Evaluating and enhancing the logical reasoning, factuality, and hallucination mitigation capabilities of Large Language Models.

Medical AI

Applying machine learning methodologies to healthcare challenges, including clinical data processing and predictive modeling.

Research Projects

Marathi-XRetrieve: Cross-Lingual Retrieval Benchmark

Independent Research

Python | BGE-M3 | ChromaDB | PyTorch | Contrastive Learning

Objective: Quantify and address performance gaps in cross-lingual retrieval for Marathi, a low-resource Indic language with over 80 million speakers.

Methodology: Developed a reproducible benchmark with 1,000 Marathi queries over English/Hindi Wikipedia. Evaluated BGE-M3 zero-shot performance and applied contrastive fine-tuning on 1,000 query-chunk pairs.

Key Results:

  • Baseline cross-lingual gap: 7.93 percentage points in Top-1 accuracy (Marathi vs. English queries)
  • Fine-tuning reduced retrieval error by 13% relative for Marathi queries (77.36% to 80.20% Top-1 accuracy)
  • Gap reduction: 47.4% with no significant degradation on English performance (-0.92 pp)
  • Comprehensive evaluation using MRR, NDCG@k, and MAP metrics with bootstrap confidence intervals
Top-1 Accuracy: Marathi Queries
Baseline
77.4%
Fine-Tuned
80.2%
13% error reduction
GitHub Repository Technical Report (Zenodo) DOI: 10.5281/zenodo.20789161

RAG-Technical-QA

Python | Retrieval-Augmented Generation | Technical Documentation

Retrieval-Augmented Generation system for answering technical questions from domain-specific documentation. Implements chunking strategies, embedding selection utilities, and evaluation frameworks for QA tasks.

View Repository

Research Experience

Undergraduate Research Assistant (Computational)

Department of Mechanical Engineering | August 2025 – December 2025
  • Model Development and Benchmarking: Deployed and systematically compared multiple machine learning frameworks—including Linear Regression, Decision Trees, Random Forests, and Artificial Neural Networks—to address a complex multi-output regression problem in manufacturing process optimization.
  • Validation Methodology: Implemented K-Fold cross-validation and rigorous hyperparameter tuning protocols to ensure robust model generalization and mitigate overfitting on real-world experimental datasets.
  • Feature Analysis: Conducted feature importance analysis on critical machining parameters (cutting speed, feed rate, depth of cut) to identify primary drivers of system performance and inform process improvements.
  • Interdisciplinary Collaboration: Partnered with doctoral researchers to translate domain-specific manufacturing parameters into actionable predictive insights, bridging mechanical engineering expertise with machine learning methodology.
Scikit-learn K-Fold Cross-Validation Hyperparameter Optimization Feature Importance Analysis Multi-output Regression

Technical Skills

Programming Languages

Python, SQL, Bash, JavaScript (foundational)

Machine Learning & NLP

Hugging Face Transformers, BGE-M3, Contrastive Learning, RAG Architectures, Evaluation Metrics (MRR, NDCG, MAP), Scikit-learn, K-Fold Cross-Validation, Feature Engineering

Models & Frameworks

Linear Regression, Decision Trees, Random Forests, Artificial Neural Networks, PyTorch, ChromaDB, FastAPI

Platforms & Tools

Git/GitHub, Kaggle, Hugging Face, Zenodo, Docker, Jupyter, GitHub Actions, Linux/Unix

Languages

English (fluent), Marathi (native), Hindi (proficient)

Contact

I welcome inquiries regarding research collaborations, Master's program opportunities, or technical discussions related to low-resource NLP, retrieval, and applied AI.

Email GitHub Kaggle Hugging Face LinkedIn Curriculum Vitae (PDF)