cv

M.Tech in Computer Science and Engineering

Student Researcher

• Developed dynamic input size calculation based on sampling rate and EEG segment duration,enhancing data flexibility.
• Automated HDF5 dataset loading for seamless integration across multiple EEG configurations.
• Integrated vector quantization training with optimized codebook and distributed data gathering for synchronized multi-node performance.
• Engineered a DeepLIFT pipeline with a custom model wrapper to compute and visualize feature attributions for MDD EEG data.
• Designed visualization tools (heatmaps, box plots, and time-resolved plots) to analyze channel-wise EEG feature importance.

Google Summer of Code (GSoC) Contributor

• Preprocessed and cleaned 5 large EEG datasets from various sources
• Customized and adapted multiple code-bases to integrate with our preprocessed EEG data
• Accelerated distributed processing and model training across multiple nodes using Pytorch and HuggingFace Trainer API, streamlining the model implementation process and reducing training time by 20%
• Developed and trained a flexible foundational EEG model, finetuned and achieved 70% accuracy on downstream tasks
• Revised multiple generative pretrained transformer architectures (GPTs), providing options to choose any based on need. Reducing Computation load by 10%
• Utilized PyTorch Captum's DeepLift and Gradient descent algorithms to understand model interpretability, and implemented model profiling for the codebase to monitor and ensure proper resource utilization

AI/ML researcher

• Directed a natural language query agent to efficiently retrieve data from a MongoDB collection utilizing Natural Language Processing and Retrieval Augmented Generation(RAG) resulting in improved user interaction and data retrieval
• Developed and designed pipelines for data preparation and testing for a novel anti-spoof detection model by combining 3 datasets totaling over 55GB, utilizing various filters to enhance data quality
• Deployed the trained model on cloud machine, granting open API access through Flask API, resulting in a 38% reduction in processing time for data analysis tasks
• Pioneered a data downloading channel in Python that efficiently retrieved and processed over 3000+ videos, reducing data acquisition time by 50%

Student Researcher

• Developed advanced text-based inpainting pipelines, enabling seamless context-aware image modifications using Stable Diffusion.
• Engineered multi-face clustering solutions with DBScan, optimizing face grouping and identity preservation in complex image datasets.
• Fine-tuned Stable Diffusion models, enhancing image generation quality, control, and consistency for diverse tasks.
• Optimized inference pipelines, implementing efficient GPU memory management and accelerated batch processing.
• Integrated custom embeddings and LoRA fine-tuning, allowing for style adaptation and domain-specific image synthesis.
• Implemented a novel multi-res OCR method to localize text from complex vector diagrams.

ML Infra/ Backend

ComicStrips (Text-to-Comic Strip Generation)

• Developed python based reddit crawler and scraped over 1,000 images, then fine-tuned the flux-dev model using Low Rank Adaption technique (LoRA) for text-to-comic strip generation
• Prepared an automatic data annotation pipeline for captioning scrapped images using Large Vision Model while data preparation
• Deployed the model on Hugging Face, achieving 20,000+ downloads, highlighting the project's reach and user interest

Adhoc (Automated Documentation Tool)

• Created a command-line tool to automate codebase documentation with local large language models, generating professional outputs in LaTeX, Markdown, and Word formats
• Released as a Python pip package, widely adopted by peers simplifying the generation of detailed, professional documentation for codebase changes with minimal effort

AnimeStudio (Anime Edits)

• Fine-tuned a Stable Diffusion model for real-time Ghibli style image transformations (1.3k+ conversions at sub-100ms median latency), and deployed a secure proxy with Upstash Redis for IP-based rate limiting, scaling to 11GB+ bandwidth globally with minimal error rates.
• Mobilized a production grade FastAPI backend with background jobs and automated cleanup, handling 10k+ daily requests in 40+ countries (∼ 140 unique users/day in the first week)

• Academic Rank 7th in the Department of Computer Science and Engineering after 5 semesters.

• First and one of the only two students selected for Google Summer of Code (GSoC) from VIT Amaravati

• Led a team of 5 international teammates to achieve a Top 50 finish in HackSqaud 2023 (Top 60 winners out of 2500+ teams).