Fine tuned open source large language models to generate genre conditioned horror stories and fairy tales, then evaluated how well different models match the tone and patterns of the training data.
Tech Stack: Falcon 7B Instruct, Llama 2 7B, Hugging Face Transformers, PEFT, LoRA, bitsandbytes, PyTorch, BERT, LDA, Python
Explored multiple deep learning architectures for automatic image captioning – starting from a CNN-RNN baseline, extending it with attention, and finally fine tuning a ViT-GPT2 encoder–decoder model – then compared their quality using ROUGE scores and test loss.
Tech Stack: PyTorch, Hugging Face Transformers, Inception v3, LSTMs, Vision Transformer (ViT), GPT-2, spaCy
Built a full camera calibration and marker-based AR pipeline using OpenCV. Estimated real-world pose from a checkerboard, computed projection transformations, and rendered a virtual 3D prism anchored to the physical marker.
Tech Stack: C++, OpenCV, solvePnP, projectPoints, Linear Algebra, 3D Geometry
Built a real-time 2D object recognition system in OpenCV that segments live video, computes custom shape descriptors, and classifies objects like mugs, gloves, watches, and power banks while allowing users to register new categories on the fly.
Tech Stack: C++, OpenCV, Hu moments, k-NN style classifier, cosine similarity, CSV feature store
Built a deep learning system that classifies MRI brain scans and highlights tumor regions using a combination of CNNs, data augmentation pipelines, and transfer learning methods. Designed for reliability across multiple datasets and MRI formats, achieving strong F1-scores and stable learning dynamics.
Tech Stack: PyTorch, Keras/TensorFlow, EfficientNetB4, YOLOv5, OpenCV, NumPy, Matplotlib