About Me

Hello! My name is Aseem Dandgaval and I’m an MS in Data Science grad at UC San Diego. I’m passionate about building impactful solutions through ML at scale. I am open to working on uncharted challenges and am always ready to learn with a positive and strong mindset. If you’re at this page, lets connect!

Blog

Mastering Memory Optimization in ML Systems: Article

Education

University of California San Diego

• Master of Science in Data Science
• GPA: 3.84/4
• Courses: Statistical Models, Probability and Statistics, Numercial Linear Algebra, Data Management for Data Sci, Responsible Data Sci, Causal Inference, Machine Learning, Visual Learning, Scalable Data Systems, ML Systems, Natural Language Processing, DataSci for Biomedicine, Data Ethics

International Institute of Information Technology, Pune

• Bachelor’s in Computer Engineering (Honors in AI-ML)
• GPA: 8.59/10
• Relevant Courses: Computational Statistics, Machine Learning, Deep Learning, Natural Language Processing, Artificial Intelligence, Business Intelligence, Data Science and Big Data Analytics, Data Structures and Algorithms.

Work Experience

CureSkin | Data Science Intern (Computer Vision)

San Diego, USA | Jun. 2024 - Sep. 2024

• Implemented multiple deep learning architectures like FCN and PSPNet using PyTorch for semantic segmentation, aiding in the localization of skin conditions such as melasma.
• Fine-tuned state-of-the-art Vision Transformers, including SwinV2, BeiT, and DeiT, for enhanced performance in image classification tasks.
• Designed custom metrics to measure annotation consistency, automating the performance tracking of various annotators.
• Conducted statistical analysis to identify the optimal annotation strategy for small skin conditions such as acne and pimples.

CureSkin | Data Science Intern (Computer Vision)

Bengaluru, India | Apr. 2022 - Jan. 2023

• Iteratively visualized data through graphs, executed custom metric calculations, and conducted comprehensive tests across multiple datasets, informing crucial data-driven decisions and assessing project health.
• Experimented with state-of-the-art approaches from academic papers for Facial Image Inpainting and generation of 3D Face Meshes for head tracking.
• Designed and implemented custom RMSE and MAE metrics for spatial model architectures to benchmark performance and determine the best fit for the data.
• Deployed models using Flask and optimized deployment, reducing time complexity by 50%.
• Automated image annotation/labeling and mask generation tasks using Python that streamlined model input pipelines.

Projects

MiniLLM | Python, Pytorch, HuggingFace, tiktoken, Gradio | (GitHub) | (Website)

• Implemented GPT2-14M from scratch and trained on a cluster of 8 × 80GB A100 GPUs on 10 Billion tokens.
• Used strategies like Mixed Precision Training, Data Distributed Training, etc to optimize training (∼ 3.5 hours).
• Acheived a lower validation loss and matched the performance of the OpenAI checkpoint on the HellaSwag eval.
  

Image Dehazing | Python, Pytorch | (GitHub)

• Implemented a Pix2Pix GAN from scratch to perform image dehazing, improving visual clarity and detail in hazy images.
• Modified the original architecture to incorporate bilinear upsampling, resulting in improved PSNR and SSIM scores, and enhancing the quality and accuracy of dehazed images.
  

Food Calorie Estimator | Python, Pytorch, YOLOv5 | (GitHub)

• Implemented an object detection model (YOLOv5) for 20 different food items with 98.9% accuracy.
• Calculated calories of food items by finding it’s mass and volume occupied in bounding cuboid using the reference object in two optical views with food items having less than 10% error with respect to ground truth.
  

Graph Network Analysis | Python, SpaCy, NetworkX | (GitHub)

• Parsed all Harry Potter books to extract interpersonal relationships of all characters using Named Entity Recognition.
• Created interactive network graphs, identified groups of characters using community detection, and tracked the importance of all characters over the entire series of books. (
  

Fraud Detection | Python, Pandas, Seaborn, Scikit-Learn, XGBoost | (GitHub)

• Cleaned, pre-processed and analyzed over 6 million rows of banking transactions.
• Revealed the most important features that help detect fraudulent transactions and tested several ML models to get an F-1 score of 76.16 on a highly skewed dataset.
  

GalleryGAN | Python, PyTorch | (GitHub)

• Developed a Generative Adversarial Network (GAN) that generates images of Abstract Art from scratch.
• Final Images of art are generated from Gaussian noise and are of size 512x512 and 8 bit color.
  

Exploratory Data Analysis on Industry Salaries | Python, Pandas, MatplotLib, Seaborn | (GitHub)

• Performed an Exploratory Data Analysis on the salaries of working professionals in the data industry in India (Data Scientists, Machine Learning Engineesrs etc).
• Cleaned and pre-processed a real world dataset using Pandas and NumPy.
• Developed visual graphs based on sub-domains, years of experience, employers and inferred patterns such as salary gap between locations, salary ranges as per experience and more.
  

Pet Breed Classifier | Python, fast.ai | (Dataset)

• Made a CNN classifier that classifies between 23 breeds of casts and dogs.
• The result is a classifier model made by transfer learning with an accuracy of 85%.
• The data set is made by me completely from scratch by web scraping and consists of 3800 image.
  

Technical Skills

• Languages/Tools: Python, SQL, PostgreSQL, C++, Git, Terminal, Linux
• Libraries/Frameworks: Pytorch, Tensorflow, HuggingFace, Scikit-Learn, Pandas, Ray, Gradio