Application of Natural Language Processing techniques to improve translation accuracy of ancient and low-resource languages. Developed classification-translation models for Sumerian-English transliteration and Named Entity Recognition models to bridge lexical gaps from unique Sumerian proper nouns, enhancing translational accuracy and semantic understanding.
Custom TF-IDF vectorization pipeline optimized for sentiment classification with frequency-based feature selection and context-aware n-gram generation (n=1,2,3). Implemented hashmap-based term frequency calculator reducing computational complexity from O(n²) to O(n). Achieved 4.2% accuracy improvement over scikit-learn's baseline on IMDb dataset, particularly excelling at identifying nuanced sentiment expressions and negation patterns.
Hybrid OLTP/OLAP data warehouse utilizing SAX parsers for efficient XML processing with O(1) memory complexity. Custom metadata-driven ETL pipeline with XPath query optimization reduced extract processing time by 68% compared to DOM parsing. Snowflake schema design with SCD Type 2 for historical tracking, materialized views for aggregation acceleration, and MySQL stored procedures for transactional consistency.
Cross-platform file system abstraction tool implementing custom B-tree indexing with O(log n) query performance. Platform-agnostic core translates between POSIX and Win32 filesystem APIs through adapter pattern, enabling unified XPath-inspired query execution across heterogeneous systems.
Custom K-Means clustering algorithm for text data analysis on the 20 Newsgroups dataset. Utilized TF-IDF vectorization to categorize text documents into clusters, enhancing understanding of content similarities. Benchmarked against scikit-learn KMeans using Homogeneity, Completeness, V-measure, Adjusted Rand Index, and Silhouette Coefficient metrics, focusing on optimizing clustering quality and execution efficiency for complex textual data.
C++ implementation of the theoretical framework for energy-aware task scheduling in mobile cloud computing. Constructs Directed Acyclic Graphs (DAGs) of tasks with local and cloud execution times. Performs initial task assignment and prioritization to minimize total completion time, then iteratively explores task migrations between local cores and cloud for energy efficiency. Linear-time kernel rescheduling algorithm ensures all precedence constraints and resource capacities remain satisfied while optimizing both makespan and energy consumption.
Simulation exploring Q-Learning, SARSA, and Actor-Critic algorithms in a complex stochastic grid-based maze environment (248 navigable cells) with dynamic challenges including walls, bumps, and oil slicks. Each algorithm evaluated over 10 independent runs of 1,000 episodes, optimizing paths and maximizing rewards within defined penalty and reward structures.
Applied reinforcement learning to fine-tune activation of the p53-Mdm2 negative feedback loop, crucial for cell cycle regulation and tumor suppression. Developed optimal control policy to effectively manage key network components (ATM, p53, Wip1, MDM2), maximizing therapeutic efficacy by maintaining an active p53 pathway. This novel approach offers potential for inhibiting tumor growth and enhancing cellular repair mechanisms in cancer therapy.
Bayesian-optimized MLP architecture for minimum-error classification of multivariate random vectors with non-Gaussian distributions. Custom regularization combining L2 weight decay with Bayesian priors to approximate posterior probability distributions, achieving 97.2% accuracy. Implemented adaptive learning rate scheduler, conditional batch normalization for non-IID data, and entropy-weighted loss function. Parameterized grid search across 128 configurations identified optimal single hidden layer with 42 neurons, achieving classification error within 0.3% of theoretical Bayes error limit.
ETL pipeline implementation for efficiently processing and transforming CSV files into structured data points and features. Responsive to API calls, the system methodically extracts data from CSV files, applies necessary transformations, and uploads processed data to a relational database. Facilitates dynamic data management, supports data analysis, and provides scope for extended API integration.
Implementation of Maximum A Posteriori (MAP) estimation to enhance vehicle localization accuracy in 2D environments. Bayesian model integrates range measurements to multiple landmarks with existing positional data, enabling highly accurate estimation of vehicle's most probable location. Effectively addresses noisy distance measurements and prior knowledge, ensuring precise navigation despite environmental complexities and measurement inaccuracies.
Game engine for the classic Snake game developed using Vue.js with vanilla HTML and CSS. Vue.js streamlines state management and facilitates real-time rendering, ensuring dynamic updates are efficiently handled without browser reloading. Implements dynamic management of game state including snake movement, collision detection, and score tracking with clean color schemes, borders, and animations.
Game engine for Tic-tac-toe developed using vanilla JavaScript, HTML, and CSS. Implements the Minimax algorithm to simulate intelligent opponent moves and strategize optimal responses. The algorithm enables the system opponent to make strategically sound decisions, mimicking human-like decision-making processes, adding challenge and realism to the game engine.
Sound processing and song note compilation platform for real-time audio signal processing and synthesis. Enables generation of various waveforms (Sine, Square, Triangle, Sawtooth) through user-adjustable parameters. Supports complex audio signal layering with detailed amplitude control and modulation. Advanced digital filters including reverb create short-lived echoes to simulate different acoustic environments. Song player functionality decodes various audio formats, arranging decoded samples into seamless playback sequence for sound design or educational purposes.
Summer 2020 Internship Project
Android companion app to interface with UV light hardware device component of Purple-Bug UV ride-share cleaning product. Developed using Java and Kotlin with stable and responsive UI. Network communications implemented using Retrofit for effective data transmission between mobile app and backend services. Image loading and caching optimized using Glide. Local data storage implemented using Room (abstraction layer over SQLite). Complex asynchronous tasks managed using RxJava and Kotlin Coroutines for smooth operation.