Projects
💡 Algorithm-to-Python Source Code Translator using LLMs
Tech Stack: Python, Hugging Face Transformers, Mistral 7B, LoRA, PEFT, Gemini API
Fine-tuned Mistral 7B using LoRA (rank=32, 4-bit quantization) to automate the translation of algorithmic descriptions into executable Python code.
- Built a custom dataset of 18,000 algorithm-code pairs using the Gemini API, publicly released on Hugging Face
- Achieved 93.5% accuracy and perplexity of 2.03, outperforming prior Seq2Seq approaches (88.7%)
- Presented at ITECHCET 2024 & published in AIP Conference Proceedings
🌊 Biofouling Detection and Classification using Deep Learning
Tech Stack: Python, PyTorch, EfficientNetV2, DenseNet121, YOLOv8, CBAM, BYOL
Developed a deep learning framework for six-class marine biofouling classification under simulated underwater imaging conditions. Manuscript in preparation.
- Built a custom underwater degradation augmentation pipeline (colour attenuation, turbidity, backscatter, non-uniform illumination) to expand a 448-image dataset to 3,443 training samples
- Evaluated seven model variants across EfficientNetV2, DenseNet121, and YOLOv8m with CBAM attention and BYOL self-supervised pretraining
- Proposed EfficientNetV2-S + CBAM + BYOL achieved validation mAP@50 of 0.9624 and macro F1-score of 0.94, suitable for deployment on autonomous underwater vehicles
🤖 Local RAG Chatbot with Document QA
Tech Stack: Python, LangChain, ChromaDB, Sentence Transformers, Gemma 2B, Ollama, FastAPI, Docker
A fully offline retrieval-augmented generation chatbot supporting context-grounded question answering over uploaded documents.
- Integrated Gemma 2B via Ollama with ChromaDB vector search and sentence-transformers for semantic retrieval
- Built a FastAPI backend supporting PDF and TXT document upload
- Deployed entirely via Docker with no external API dependencies
🧠 Deep Learning Implementations from Scratch
Tech Stack: Python, PyTorch, CIFAR-10, MNIST
Implemented core deep learning architectures from scratch to build a strong foundational understanding of how modern models work at every level.
- Built MLP, CNN, VGG19, DenseNet, and InceptionV3 on CIFAR-10 and MNIST
- Implemented an RNN for Netflix stock price forecasting
💼 Feature Selection & Salary Prediction using Rough Set Theory
Tech Stack: Python, Scikit-learn, Pandas
Applied Fuzzy Rough Set Theory with the Quick Reduct algorithm for feature selection on Stack Overflow Developer Survey data (2018–2020).
- Identified the minimal subset of attributes most predictive of salary (education, location, years of experience)
- Trained a Decision Tree regression model achieving R² scores of up to 0.9997 and MAE as low as 153.32
🌫 Air Quality Prediction Model (PM2.5)
Tech Stack: Python, Scikit-learn, Pandas
Built a regression model to predict PM2.5 pollutant levels from 36,000+ hourly air quality records (2017–2022).
- Engineered temporal features from raw timestamps and applied Label Encoding for categorical time bins
- Compared Linear Regression, Decision Tree, and Random Forest regressors with GridSearchCV hyperparameter tuning
- Achieved RMSE of 8.97 with the optimized Decision Tree model
🌐 Personal Portfolio Website
Tech Stack: HTML, CSS, JavaScript
A responsive portfolio site showcasing projects, publications, and contact information.
- Custom dark/light modes & animated transitions
- Clean layout with accessible design