Current Projects

...

Previous Projects

My GitHub contributions:

• Visual Question Answering with Generative AI
  Integrated Hugging Face pre-trained tokenizers, Visual Transformer for images, and LLMs for generating answers. Achieved 0.3 WUPS with RoBERTa and BEiT outperforming all 4 model combinations viz. ViT, DEiT with BERT, GPT. Deployed multimodal VQA in Docker for containerization, Kubernetes for orchestration, ensuring scalable, efficient service.
• DS5983: Large Language Models
  
  • Project 1 (N-gram models): Test various n-gram models on Reuters.
  • Project 2 (Transformer Architecture Implementation and Machine Translation): Implemented Transformer model architecture form scratch. Train the model on a machine translation task from German to English using the Multi30k dataset. Implemented greedy decoding and beam search for translation.
  • Project 3 (Text Summarization with HuggingFace BART): Evaluated summarization of 3 pre-trained Hugging Face Transformer models BART, T5, Pegasus on SAMSum dataset. Improved BART model’s ROUGE score on testing set from 28.7 to 37.5 by fine-tuning with Mixed Precision (AMP).
  • Project 4 (Prompt Engineering and Few-Shot Learning with Flan-T5 for Dialogue Summarization): Performed prompt engineering on Flan-T5 using dialogsum dataset with instructional prompts and pre-built T5 prompts. Experimented with zero-shot and few-shot learning to assess their impact on summaries' relevance and coherence.
  • Project 5 (Fine-Tuning Flan-T5 with PEFT (LoRA)): Fine-tuned Flan-T5 LLM for dialogue summarization, employing full and parameter-efficient techniques (PEFT), achieving enhanced performance metrics evaluated via ROUGE scores. Applied Low-Rank Adaptation (LoRA) to optimize training efficiency, reducing computational resources while maintaining high accuracy in AI-driven summarization tasks.
• Camera Calibration and Augmented Reality Integration with Checkerboard
  Applied calibration techniques with checkerboard, camera pose estimation to align virtual objects within AR environment. Explored feature detection (Harris corners, SURF) for target recognition, and overlayed 3D virtual objects onto 2D video.
• Image colorization of historical paintings with GAN
  Leveraged U-Net and pix2pix Convolutional Generative Adversarial Network, to colorize grayscale historical paintings. Utilized CIELAB color space conversion and Patch discriminator for enhanced image-to-image translation.
• Pothole Detection and Segmentation Jan 2020 - Jun 2020
  Fine-tuned custom Mask R-CNN and YOLACT instance segmentation models for real-time pothole detection and segmentation on Indian roads with PyTorch and achieved 86% accuracy, 0.30 mAP on custom testing dataset
• Udacity's Deep Analyst Nanodegree projects Apr 2020 - Jun 2020
  
  • Project 1: Explore Weather trends analyzes local and global temperature data and compare the temperature trends where you live to overall global temperature trends.
  • Project 2: Investigate a dataset analyzes a dataset and then communicates the findings.
  • Project 3: Analyze A/B test results understands the results of an A/B test run by an e-commerce website.
  • Project 4: Wrangle and Analyze Data wrangles and analyzes the tweet data of WeRateDogs that rates people's dogs with a humorous comment about the dog.
  • Project 5: Communicate Data Findings deals with data exploration of the flights cancellation and delay dataset mainly using data visualization and present the findings using explanatory visualizations.
• Udacity's Deep Reinforcement Learning Nanodegree projects in PyTorch Jul 2019 - Nov 2019
  
  • Project 1: Navigation is about training a RL agent to navigate (and collect bananas!) in a large, square world.
  • Project 2: Continuous Control is about training a RL double-jointed arm agent so that it can move to target locations.
  • Project 3: Collaboration and Competition is about training two RL agents to control tennis rackets to bounce a ball over a net.
• Udacity's Deep Learning Nanodegree projects in PyTorch Dec 2018 - Mar 2019
  
  • Classification: Flower image classification and Dog breed classification
  • Generation: TV Script generation and Face Generation using DCGAN
  • Deployemnt: Deploying sentiment analysis model on Sagemaker
  • Art: Neural Style Transfer
• ImageCaptioner: Image captioning using Encoder-Decoder Jan 2019 - Feb 2019
  Developed image captioning application based on Neural Image Caption model utilizing encoder-decoder architecture, using pretrained CNN as encoder and LSTM as decoder.
• Plant Disease Detection and Recognition Aug 2018 - Sep 2018
  Used Transfer Learning to develop a plant disease detection and recognition system. Applied ResNet101 architecture of Convolutional Neural Networks in the model. Developed an Android app to display the system
• TweetSense: Real-time social media sentiment analysis July 2018 - Aug 2018
  Developed an application that analyzes tweets and intelligently provides real-time feedback, using sentiment analysis, in a visual manner with the help of a time series graph. The app also provides the sentiment analysis of tweets in the last week. TweetSense also analyzes user's text and provides tone analysis (using Watson API) in addition to sentiment score.
• Railway ticketing system Apr 2018
  Developed a desktop app in Qt (C++ cross-platform framework) for railway ticketing system.
• DengueApp: Location based dengue prediction Jan 2018 - Mar 2018
  Developed an Android app that gives real-time location-based dengue risk index to the user using machine learning techniques. The model used the weather conditions of the user’s location as features. The gradient boosted trees performed better than other algorithms. Hosted the model on Django server as an API. Showed the application at Smart India Hackathon 2018 grand finale.
• Housing price prediction Jan 2018
  Predicted housing price using regression techniques. Applied different models (linear regression, Lasso, Ridge, boosting, random forest). xgboost outperformed other models.
• Titanic survivors Oct 2017
  Predicted survival on the Titanic using machine learning techniques. Tried different models (SVM, kNN, logistic regression, random forest); random forest gave the highest accuracy.
• Technical Fridays - personal website and blog July 2017
  This is my personal website and blog developed with HTML5, CSS and JavaScript using Jekyll and Github pages.
• Blog in django March 2017
  A blog developed in Django, a high-level python web framework using PostgreSQL database.
• python-projects Feb 2017
  It is a series of small programs written in python.