Projects & Publications

A showcase of my technical projects, research work, and contributions to the field of computer science and software engineering.

Publications

Automatic Generation of Math Word Problems for Assessing Learner Skills in Adaptive Learning Systems

Publication

Lavanika Srinivasaraghavan

February 2022, Hyderabad, India

Generated algebra math word problems using neural networks for enhancing intelligent tutoring in adaptive learning systems, reducing manual effort by 40%. Achieved a 60% increase in BLEU scores using transformer-based LLM models GPT-2 and TFGPT2LMHead (with Char-LSTM as a baseline), enhancing problem generation accuracy and quality.

GPT-2 TFGPT2LMHead Char-LSTM Neural Networks Adaptive Learning

More Projects

Sentiment Analysis in Dravidian Languages

Detected homophobic and transphobic content in over 3,000 Tamil-English YouTube comments as part of Dravidian LangTech Workshop 2022.

BiLSTM NLP ML Ensemble models

Sentiment Analysis in Dravidian Languages

Detected homophobic and transphobic content in over 3,000 Tamil-English YouTube comments as part of Dravidian LangTech Workshop 2022. Conducted comparative study of ML models, ensemble models, and BiLSTM, with BiLSTM achieving the highest macro F1 score. The project involved preprocessing code-mixed text, handling Tamil script and transliterations, and developing robust models to identify harmful content in social media discussions. This work was part of a shared task focused on improving content moderation for underrepresented languages.

Multi-Class Image Classification

Designed custom CNN model in PyTorch for six-class malaria cell classification (5,000+ images) achieving 95% test accuracy.

PyTorch CNN ResNet-50 Transfer Learning

Multi-Class Image Classification

Designed custom CNN model in PyTorch for six-class malaria cell classification (5,000+ images) achieving 95% test accuracy. Adapted ResNet-50 for transfer learning approach obtaining 92% accuracy. Optimized hyperparameters, reducing training loss by 80% for CNN and 58% for ResNet-50. The project involved extensive data augmentation techniques, batch normalization, and dropout regularization to prevent overfitting. Conducted detailed ablation studies comparing different architectures and training strategies. The high accuracy on unseen test data demonstrates the model's strong generalization capabilities for medical image analysis.

Information Retrieval Projects

Implemented the full modern IR stack from scratch: evaluation metrics, inverted indexes with compression, classical ranking models (VSM, BM25, QL), RM1/RM3 feedback, and more.

TF-IDF Ranking models Web search engines NLP

Information Retrieval Projects

Implemented the full modern IR stack from scratch: evaluation metrics, inverted indexes with compression, classical ranking models (VSM, BM25, QL), RM1/RM3 feedback, TextRank, TF-IDF K-means clustering, LambdaMART, and a dense dual-encoder with FAISS. Built efficient indexing structures using variable byte encoding and delta encoding for compression. Developed query expansion techniques using pseudo-relevance feedback. Implemented learning-to-rank models using gradient boosting. Created neural retrieval systems using BERT-based dense encoders and approximate nearest neighbor search with FAISS for fast similarity matching at scale.

Artificial Intelligence Projects

Built a Messenger Bot using AIML for keyword-based intent detection and a Career Expert System using Experta for personalized recommendations.

AIML Messenger Bot Experta Expert System OpenWeatherMap API

Artificial Intelligence Projects

Built a Messenger Bot using AIML for keyword-based intent detection to recommend food outlets and provide real-time weather updates via OpenWeatherMap API. The bot uses natural language processing patterns to understand user queries and responds with contextually relevant information. Developed a Career Expert System using the Experta Python library to generate personalized course and career recommendations based on GPA, publications, and achievements. The expert system uses rule-based reasoning to match student profiles with optimal career paths, considering factors like academic performance, research interests, and extracurricular activities.

Classification of Math Word Problems based on Difficulty

Performed multiclass classification of math word problems into different difficulty levels using Readability score, Math vocabulary, and problem length.

Neural Networks Keras Multiclass Classification NLP

Classification of Math Word Problems based on Difficulty

Performed multiclass classification of math word problems into different difficulty levels (easy, medium, hard) using three attributes: Readability score, Math vocabulary, and problem length. Utilized Neural Networks (Keras Sequential model) with three dense layers, tanh and softmax activation, and Adam optimizer. Feature engineering involved computing readability scores, extracting mathematical terminology frequency, and normalizing problem lengths. The model achieved strong performance in distinguishing between difficulty levels, demonstrating the effectiveness of combining linguistic and mathematical features for automated problem difficulty assessment.

Mutual Book Exchange Portal

Built a web application using Spring Boot, Thymeleaf, and H2 database enabling registered users to exchange books for temporary use and return.

Spring Boot Thymeleaf H2 Database Java OOP

Mutual Book Exchange Portal

Built a web application using Spring Boot, Thymeleaf, and H2 database enabling registered users to exchange books for temporary use and return, with admin functionality as a 'super-user'. Designed 3 main features: publish/view/search book details for lending and borrowing, manage wallet transactions, and enforce penalties for late return. Implemented RESTful APIs for CRUD operations, JWT-based authentication, and role-based access control.

Microsoft Engage - Tic Tac Toe Game

Developed a Tic Tac Toe Game Web application using Minimax algorithm powered by AI with feature for playing with a friend.

JavaScript Node.js Bootstrap Minimax Algorithm Azure

Microsoft Engage - Tic Tac Toe Game

Developed a Tic Tac Toe Game Web application using Minimax algorithm powered by Artificial Intelligence with feature for playing with a friend. Used Javascript, Node.js, Bootstrap and deployed using Microsoft Azure. The AI uses the Minimax algorithm with alpha-beta pruning to calculate optimal moves, making it unbeatable when playing perfectly. Implemented two game modes: Player vs AI and Player vs Player. The application features a clean, responsive interface built with Bootstrap and real-time game state updates. Part of Microsoft Engage Mentorship Program 2020.

Python Game Development - Tetris

Built the Tetris game using pygame library in a 20-day workshop organized by The Girl Code.

Python Pygame Game Development

Python Game Development - Tetris

Built the Tetris game using pygame library in a 20-day workshop organized by The Girl Code, a non-profit organization dedicated to empowering girls in STEM. Implemented core game mechanics including rotation with wall kicks, collision detection and line clearing with scoring system. Added features like next piece preview, score tracking, smooth animations and background music. The project demonstrated fundamental game development concepts including game loops, event processing, and state management using Object-Oriented Programming principles.

Interested in Collaboration?

I'm always open to discussing new projects, research opportunities, or technical challenges.