Pranjali Chaudhari

About

Who I Am

I'm a M.Tech Data Science student at MMCOE, Pune (CGPA 8.61/10), with a unique blend of a B.Tech Mechanical Engineering foundation and advanced expertise in AI/ML engineering. This cross-domain background gives me an analytical edge in approaching complex problems.

My research sits at the intersection of AI Security and Medical AI. I investigate adversarial vulnerabilities in Large Language Models and build multi-layer defense frameworks, while simultaneously exploring explainable deep learning for cervical cancer detection.

I've had hands-on industry experience at DATASMITH AI Solutions, where I built end-to-end ML pipelines for real-world datasets. I believe in writing clean, well-documented code and in making AI not just powerful — but interpretable and trustworthy.

When not training models, I'm reading research papers, experimenting with RAG systems, or diving into federated learning for privacy-preserving ML.

Pune, Maharashtra, India

pranjalichaudhari0408@gmail.com

linkedin.com/in/pranjali-a-chaudhari

github.com/Pranjali0408

MMCOE, Pune — M.Tech Data Science

2 Conference Papers Accepted • 1 Under Review

Technical Skills

What I Work With

Languages & Core

PythonSQL JavaJavaScript

Used in: All projects. Python is the primary language across every ML/AI project, data pipeline, and research experiment.

Machine Learning & Deep Learning

CNNResNet/VGG Transfer LearningSVM Scikit-learnXGBoost

Used in: Cervical Cancer Detection (CNN/ResNet), Crop Disease Prediction (Federated ML), Packaging Data Analysis (EDA + classical ML).

LLMs & NLP

LLMsRAG Prompt Engineering FLAN-T5Phi-3 Mini TransformersNLP

Used in: Bhagavad Gita Chatbot (RAG + LLM), ICAIS 2026 Paper (FLAN-T5 adversarial evaluation), ARML-Defense paper (FLAN-T5 + Phi-3 Mini).

AI Safety & XAI

AI SafetyAdversarial ML Grad-CAMSHAP LIMEExplainable AI

Used in: LLM defense papers (adversarial robustness evaluation), Cervical Cancer project (Grad-CAM for prediction interpretability).

Data Engineering

PandasNumPy EDAFeature Engineering SMOTEPower BI

Used in: DATASMITH internship (data pipelines), Packaging Industry Analysis (Power BI dashboards), all research datasets (preprocessing & augmentation).

Tools & Frameworks

PyTorchGit JupyterGoogle Colab VS CodeFederated Learning

Used in: PyTorch for CNN training (Cervical Cancer), Federated Learning framework (Crop Disease project), Git across all development projects.

Python

PyTorch

HuggingFace

Power BI

GitHub

Scikit-learn

SQL

Colab

XAI

Federated Learning

Experience

Industry Work

AI / Data Science Intern

DATASMITH AI Solutions, Pune

Aug 2025 – Nov 2025 | Full-time Internship

Architected and deployed end-to-end data preprocessing & cleaning pipelines for structured real-world industrial datasets, reducing downstream model errors by standardizing inconsistent data formats and handling missing values systematically.
Supported ML model training, validation, and performance evaluation workflows — including cross-validation, hyperparameter tuning, and performance benchmarking across classification and regression tasks.
Executed advanced feature engineering using Pandas and NumPy — including polynomial features, interaction terms, and mutual-information-based feature selection — leading to measurable improvement in model performance metrics (precision, recall, F1).
Collaborated in an agile team environment, writing modular and reusable Python code, maintaining version control via Git, and presenting findings through clear visualisations and reports.

Python Pandas NumPy Scikit-learn Feature Engineering ML Pipelines Git

Projects

What I've Built

2025–2026

Adversarial LLM Security — ICAIS & ARML Papers

A two-phase research project on securing instruction-tuned LLMs. Phase 1 (ICAIS 2026): Evaluated FLAN-T5-base against 120 adversarial prompts across 4 attack types — instruction override, role-play manipulation, contextual injection, and logical confusion. Introduced a lightweight lexical sanitization mechanism reducing adversarial attack success from 23.33% → 15.83% (a relative drop of 32.14%) without model retraining. Phase 2 (ARML): Proposed a novel 4-layer adaptive framework with risk formula R = αK + βS + γC, achieving 93.3% defense accuracy and F1 = 0.94 with only 7.7% false positive rate across FLAN-T5 and Phi-3 Mini.

Python FLAN-T5 Phi-3 Mini HuggingFace AI Safety NLP JailbreakBench

32.14%

Relative Attack Reduction (Phase 1)

93.3%

Defense Accuracy (ARML)

0.94

F1-Score (ARML)

7.7%

False Positive Rate

2025–Present

Cervical Cancer Detection — Multimodal Deep Learning

Building a CNN-based model for cervical cell classification (SIPaKMeD dataset) integrated with clinical risk features (demographics, HPV status). Applying Grad-CAM and LIME for Explainable AI to generate clinically interpretable saliency maps. Synthesised 20+ IEEE papers on multimodal DL and XAI for cervical cancer — accepted and presented at ETFI 2026, DESPU.

CNN PyTorch Grad-CAM XAI SIPaKMeD Transfer Learning

2025

Bhagavad Gita & Chanakya Niti — Domain-Specific AI Chatbot

Built a domain-specific conversational AI using Retrieval-Augmented Generation (RAG). The system embeds and indexes philosophical texts, then retrieves contextually relevant passages before generating responses — dramatically reducing hallucination compared to vanilla LLM inference. Implemented chunking strategies, vector similarity search, and curated knowledge grounding to preserve authentic philosophical meaning across Gita's 18 chapters and Chanakya's Arthashastra excerpts.

LLMs RAG Prompt Engineering Vector DB Python

2026

Federated Learning for Privacy-Preserving Crop Disease Prediction

Designed a distributed ML workflow for agricultural disease prediction — trained models locally on individual farm nodes without centralising raw data. Studied and compared convergence behaviour (FedAvg vs FedProx), communication overhead, and non-IID data challenges. Implemented differential privacy mechanisms and analysed the privacy-accuracy trade-off, achieving competitive accuracy while ensuring farm data never leaves local nodes.

Federated Learning Python Differential Privacy FedAvg PyTorch

2023

Packaging Industry Data Analysis & KPI Visualisation

Cleaned and transformed production datasets from a packaging company using Pandas — resolved inconsistent formats, imputed missing values, and standardised KPI definitions. Built interactive Power BI dashboards tracking production throughput, defect rates (%), yield loss, and machine downtime. Enabled data-driven decisions reducing reporting time from days to real-time visibility.

Power BI EDA Pandas Data Visualisation KPI Design

Research

Publications & Papers

Evaluating Security and Robustness of Instruction-Tuned Large Language Models Against Adversarial Prompt Attacks

ICAIS 2026 — NMIMS University, Indore | Published: Springer LNCS

Conducted a controlled empirical assessment of adversarial robustness in instruction-tuned LLMs using 120 balanced prompts across 4 attack categories. Introduced a lightweight lexical sanitization defense reducing attack success rate from 23.33% to 15.83% — a relative reduction of 32.14% — without requiring model retraining on FLAN-T5-base.

Camera-Ready Accepted Springer LNCS → Scopus Indexed+Web of Science LLM Security Adversarial ML AI Safety

Adaptive Risk-Aware Multi-Layer Defense Framework for Securing Large Language Models Against Adversarial Prompt Injection Attacks

International Conference — Jamia Millia Islamia, New Delhi | Decision Pending

Proposed ARML-Defense — a 4-layer adaptive security framework using composite risk scoring R = αK + βS + γC with a novel Allow/Modify/Restrict policy. Achieved 93.3% defense accuracy and F1 = 0.94 on FLAN-T5 and Phi-3 Mini across 100 adversarial prompts from JailbreakBench and AdvBench. Outperforms Hard Restriction by 26.6 percentage points while maintaining 92.3% usability preservation.

Under Review — Decision Pending Multi-Layer Defense Semantic Analysis Adaptive AI

A Comprehensive Review on Multimodal and Explainable Deep Learning Approaches for Cervical Cancer Detection

ETFI 2026 — DESPU, Pune | Published & Presented (IEEE Proceedings)

Systematic review and analysis of 20+ IEEE research papers on CNN-based architectures, hybrid ML-DL models, Vision Transformers, and XAI methods (Grad-CAM, SHAP, LIME) for cervical cancer detection. Identified research gaps in multimodal integration and post-hoc explainability. Presented at ETFI 2026, Department of Electronics, Telecommunication and Future Innovations conference.

Accepted & Presented IEEE Xplore → Scopus + Web of Science (Clarivate) Medical AI Explainable AI Deep Learning

Publication Indexing Note

ICAIS 2026 (Springer LNCS): Springer Lecture Notes in Computer Science (LNCS) series is indexed in Web of Science + Scopus (Elsevier) and DBLP. | ETFI 2026 (IEEE): IEEE conference proceedings are published on IEEE Xplore, indexed in both Scopus and Web of Science (Clarivate / SCI/ESCI), subject to conference-specific inclusion. | ARML (Jamia Millia Islamia): Decision pending.

Education

Academic Journey

M.Tech in Data Science (Information Technology)

Marathwada Mitra Mandal's College of Engineering (MMCOE), Pune

2024 – Present

CGPA: 9.56 / 10

Focus: Machine Learning, Deep Learning, LLM Security, Federated Learning, Research Methodology. Active researcher with 2 accepted international conference papers during this program.

B.Tech in Mechanical Engineering

Vishwakarma Institute of Information Technology (VIIT), Pune

2019 – 2023

CGPA: 8.51 / 10

Strong foundation in analytical thinking, systems design, and engineering mathematics. Transitioned to Data Science leveraging engineering problem-solving skills.

HSC — Class XII (Science)

Maharashtra State Board

2019

84.92%

SSC — Class X

Maharashtra State Board

2017

93.20%

Credentials

Certifications & Achievements

Generative AI Fundamentals — Databricks

Python for Data Science & ML — Udemy

Python for Everybody — Coursera / University of Michigan

Full Stack Web Dev Bootcamp — Udemy

2 International Conference Papers Accepted at ICAIS 2026 (Springer) and ETFI 2026 (IEEE) during M.Tech studies

98 / 100 in Cyber Security coursework — demonstrating strong foundation in security fundamentals aligned with LLM research

32.14% Adversarial Attack Reduction achieved via novel lightweight defense — published, peer-reviewed finding