showmick

implementing the transformer architecture from scratch in pytorch

• •achieved a 14.8 BLEU score (~35% of original paper's performance) on English-Italian translation whilst using resource efficient training with single A100 GPU versus paper's 8x P100 GPU setup
• •processed 1.25M English-Italian sentence pairs (19.5M tokens) from OPUS Books dataset with custom tokenization and efficient data loading pipeline
• •deployed scalable training pipeline with checkpoint management, and Weights & Biases experiment tracking, optimizing for GPU memory constraints

most downloaded code llama adapter on the hugging face hub

• •developed an automated GitHub mining pipeline that processed over 120 FastAPI repositories, extracting production-grade code patterns—including authentication, database integration (SQLAlchemy, MongoDB), input validation, and error handling—via advanced AST parsing and pattern recognition, resulting in a comprehensive, curated dataset for model fine-tuning
• •boosted FastAPI code generation accuracy and completeness by 23.7%, leveraging custom domain-specific evaluation metrics and a GPT-based “LLM as a Judge”
• •reduced GPU memory usage by approximately 75% and enabled single-GPU (T4, 15GB VRAM) training by adopting 4-bit quantization and LoRA adapters, cutting training time by over 50% compared to full-parameter fine-tuning

implemented a single-layer perceptron (SLP) in java

• •predicts LLM affinity based on an individual's personality traits, achieving ~76% accuracy
• •supports custom-built activation functions like ReLU, ELU and sigmoid
• •follows encapsulated class structure, enums, method overloading and other object-oriented principles
• •implemented a CI pipeline with GitHub Actions to automate Java builds and test execution with JUnit

an AI assistant for archaeologists

• •built a machine learning model that analyzes historical excavation data and geospatial patterns to predict fossil-rich excavation hotspots
• •designed an AI system that generates structured excavation plans, estimating costs, workforce, logistics, and excavation duration to assist archaeologists
• •fine-tuned MobileNetV2 on 100,000+ synthetic data points and 5,000 real fossil images to classify fossils, determine species, and predict market value
• •processed and visualized over 240,000+ excavation-related data points on a world map, enabling archaeologists to explore fossil distribution interactively

a safety-focused web application for Atlanta residents

• •features real-time alerts that notify users when approaching high-risk areas and an interactive heatmap for visualizing crime data from Atlanta PD
• •incorporated a custom shape-drawing tool allowing users to select specific map areas for detailed crime analysis and an SOS button for immediate emergency calls
• •built with React.js for a responsive user interface and used MongoDB Atlas to securely store and manage historical crime data
• •implemented robust server-side communication protocols with Node.js to handle real-time data updates

implementing merge sort, insertion sort and binary search from scratch in Java

• •using the algorithms in a system to rank programming languages based on various attributes
• •utilized inheritance, polymorphism, encapsulation, amongst other object-oriented principles
• •implemented my first CI pipeline with GitHub Actions to automate Java builds and test execution with JUnit

a time-series based sustainability analyzer

• •built a dashboard to evaluate automotive sustainability by processing over 320,000 emissions records with Pandas and visualizing CO₂ trends using Matplotlib
• •trained and evaluated Random Forest and Prophet models, using feature engineering to forecast sustainability scores from vehicle emissions data
• •used OpenAI’s CLIP model for car image classification, integrated the Cohere API, and applied a fine-tuned DistilBERT model with PDFPlumber to extract insights from sustainability reports and Reddit sentiment

an auto-investing platform using AI to personalize financial planning

• •contributed by integrating the OpenAI API into the Spring Boot backend and streamlining prompt engineering for personalized investment roadmap generation
• •built and connected a Streamlit chatbot interface to enable users to interactively receive AI-driven financial plans based on their submitted financial data

developed a marketing website for a georgia tech rso (registered student organization)

• •secured 4 new sponsorships and drove a 28% increase in member engagement and recruitment

a platform to connect Georgia Tech students by class schedules

• •contributed by designing the About Me page, building key components of the Home page, and outlining the backend API structure

an AI-powered weather analytics and chatbot platform

• •developed a Streamlit-based application to analyze and visualize weather data with pandas, matplotlib, and plotly
• •offered dynamic forecasting and interactive charts powered by real-time OpenWeather API data
• •integrated Gemini API for natural language assessment, enabling users to query weather trends and receive conversational insights