Motivation: As golf courses closed and simulators became the only alternative, I noticed a steep rise in simulator prices — some charging over $90/hour. The core issues were manpower and rental overhead. To solve this, I created AnyBirdie: a fully automated experience from booking to play, powered by software and hardware integrations that manage door access, lights, air-conditioning, and simulator systems — all without human staff. This redefines accessibility to golf by making it seamless and cost-effective.

Motivation: Many people lack basic knowledge or access to legacy planning, and financial advisors often waste time copy-pasting clauses from templates. I built Absolute Legacy to democratize access to estate planning by offering professionally written legal clauses, educational guidance, and a seamless drafting experience. The goal is to shift advisors' focus from manual drafting to helping clients craft meaningful legacy solutions.

Motivation: TeleLogger was developed while building school project apps. I often worked on-the-go using my tablet or phone, and found it frustrating to access logs through command line tools. This tool was my solution for real-time, accessible logs from anywhere.

Motivation: TeleFetcher came from a Product Management project assignment to create a chatbot/agent that could summarize Telegram chats and suggest actions. I later used it in a bot for my friend group to help us plan outings, track conversations, and split bills.

Motivation: Heartbeat was built to ensure all my running tools and bots stayed live. I wanted a simple way to monitor uptime and get notified if something crashed — so I built a heartbeat service for my own projects.

Motivation: I often host gatherings with friends and wanted a more personal, elegant way to invite them — beyond plain messages or group chats. This inspired me to create Invite Service: a custom tool that integrates with my Google Calendar to display event details on a branded web page, complete with RSVP support and calendar integration.

Research Focus: This project investigates various popular compression algorithms and data structures to achieve high compression rates approaching industry-standard solutions. The research explores multiple approaches including Huffman encoding for entropy compression, and advanced heap data structures such as BinaryHeap, FourWayHeap, and PairingHeap to optimize the encoding process. Each data structure offers different tradeoffs in terms of insertion/deletion complexity, memory overhead, and encoding efficiency, allowing for comprehensive performance analysis and optimization strategies.

Technical Approach: The implementation leverages quadtree-based spatial decomposition for image segmentation, combined with sophisticated encoding techniques. By systematically comparing different heap implementations and their impact on Huffman tree construction, the research identifies optimal configurations for various image types and compression targets. The study benchmarks compression ratios, encoding/decoding speeds, and memory usage against established market solutions to validate the effectiveness of the approach.

Research Overview: Rerankers are critical components in modern information retrieval systems, operating as the second stage in a two-stage retrieval pipeline that balances efficiency with accuracy. While first-stage retrieval methods such as keyword search and vector similarity are fast and scalable, they have inherent limitations in understanding complex semantic relationships, contextual nuances, and fine-grained relevance distinctions. Rerankers address these limitations by applying more sophisticated models to a smaller, manageable set of candidates (typically 20-100 documents) retrieved by the first stage.

Experimental Design: The research conducted two comprehensive experiments. Experiment 1 established baseline performance characteristics using a controlled benchmark with 100 curated sentences and 10 carefully designed queries with ground truth relevance labels, testing four reranker providers: Cohere's rerank-english-v3.0, Voyage AI's rerank-2, Contextual AI's ctxl-rerank-v2-instruct-multilingual, and Jina AI's jina-reranker-v2-base-multilingual. Experiment 2 evaluated reranker performance in a production-like RAG system using real-world documentation chunks from Sentinel's RAGFlow instance and 19 challenging edge case queries designed to test time-sensitivity, context-dependency, negative queries, distinction queries, exception queries, role-based queries, and multi-hop reasoning scenarios.

Key Findings: The experiments revealed that Voyage AI emerged as the overall winner, achieving excellent ranking quality (MRR: 0.944, NDCG@10: 0.914), fastest latency (673ms mean), lowest cost ($0.50/1K docs), and 94.74% success rate. Cohere demonstrated the highest MRR (0.972) and excellent NDCG@10 (0.911) with good latency (877ms), making it an excellent alternative for applications prioritizing absolute best ranking quality. The research also revealed that different rerankers capture different aspects of relevance, as evidenced by low to moderate inter-reranker agreement scores, suggesting that ensemble approaches could potentially outperform individual rerankers.

Project Overview: Sentinel is an AI-powered access remediation agent designed to help employees navigate complex organizational policies and procedures for requesting access to various systems and resources. The system implements a sophisticated Retrieval-Augmented Generation (RAG) architecture that retrieves relevant documentation from a knowledge base containing company policies, user guides, system-specific documentation, and governance requirements.

Technical Architecture: The system employs a two-stage retrieval pipeline combining weighted keyword search and vector similarity for first-stage retrieval, followed by advanced reranking techniques to ensure the most relevant documents are surfaced. The research component evaluated multiple reranker solutions including Cohere, Voyage AI, Contextual AI, and Jina AI, with comprehensive benchmarking on both controlled datasets and real-world edge case queries. The system handles complex queries requiring understanding of intent, context, and relationships between concepts, ensuring accurate guidance and reducing the risk of security misconfigurations.

Architecture & Design: Tixar leverages a microservices architecture to achieve horizontal scalability and fault tolerance. The system is decomposed into independent, loosely-coupled services including user management, ticket inventory, payment processing, order management, and notification services. Each microservice can be scaled independently based on demand, allowing the system to handle traffic spikes during high-demand ticket releases.

Cloud-Native Strategies: The platform implements cloud-native design patterns including containerization with Docker, orchestration with Kubernetes, service mesh for inter-service communication, and distributed caching for performance optimization. The architecture employs event-driven patterns for asynchronous processing, message queues for reliable service communication, and distributed databases for data consistency across services. Load balancing, auto-scaling, and circuit breakers ensure high availability and resilience under varying load conditions.

Real-Time Architecture: Dancesport implements a sophisticated microservices architecture optimized for real-time data processing and live updates. The system processes judge scores, calculates rankings, and broadcasts updates to multiple clients simultaneously with minimal latency. Each component is designed for high-throughput, low-latency operations critical for live competition scenarios where timing and accuracy are paramount.

Technical Implementation: The platform leverages Server-Sent Events (SSE) for efficient one-way streaming of real-time updates from server to clients, enabling judges and spectators to receive live score updates without polling overhead. Advanced Message Queuing Protocol (AMQP) ensures reliable, asynchronous message delivery between microservices, handling score aggregation, ranking calculations, and notification distribution. The architecture implements event sourcing patterns for auditability, CQRS (Command Query Responsibility Segregation) for optimized read/write operations, and distributed consensus algorithms for maintaining consistent rankings across all system nodes. The system handles concurrent scoring from multiple judges, real-time leaderboard updates, and ensures data consistency across distributed components while maintaining sub-second response times.

Project Overview: ChatsTodo is an intelligent chat assistant that transforms conversations across multiple channels into actionable insights. The system leverages AI to automatically extract key information, create structured todo lists, and generate calendar events, seamlessly integrating with Google Suite for a unified productivity experience.

Technical Capabilities: Built as part of a Product Management assignment, ChatsTodo demonstrates advanced natural language processing capabilities for summarizing chats, identifying action items, and recommending next steps. The system processes conversations from various chat platforms, applies intelligent extraction algorithms to identify tasks, deadlines, and important information, and automatically creates corresponding entries in Google Calendar and task management systems. The platform showcases how AI can bridge the gap between communication and productivity tools, reducing manual effort and ensuring important information is never lost in conversation threads.

Research Overview: Reactive Search Optimisation represents a paradigm shift in metaheuristics, moving beyond static algorithmic approaches to create adaptive, learning-based optimization strategies. Unlike traditional metaheuristics that follow fixed rules, RSO employs "learning on the job" principles, dynamically adjusting search strategies based on feedback and performance metrics. The framework emphasizes diversity, rapid generation and analysis, and problem-specific adaptation to enhance both efficiency and effectiveness of optimization algorithms.

Reactive Strategies Explored: The research investigated four primary reactive schemes: (1) Reacting on Neighborhood - using Variable Neighborhood Descent (VND), Iterated Local Search (ILS), and Variable Neighborhood Search (VNS) to explore different solution spaces and escape local minima; (2) Reacting on Annealing Schedule - implementing adaptive and non-monotonic cooling schedules in Simulated Annealing to balance exploration and exploitation; (3) Reacting on Objective Function - applying clause-weighting and breakout methods to modify constraint weights and guide search away from explored regions; (4) Reactive Population-Based Methods - leveraging genetic algorithms, memetic algorithms, and collaborative RSO (CORSO) for knowledge sharing and multi-stream optimization.

Applications & Case Studies: The research examined real-world applications across multiple domains including classic combinatorial tasks (0/1 Multidemand Multidimensional Knapsack Problem, TSP, Vehicle Routing Problem with electric vans considering charging time and traffic), satisfiability problems (software engineer project assignment with multiple constraints), neural network optimization (hyperparameter tuning and avoiding local minima in backpropagation), and continuous optimization problems. The study demonstrated how RSO techniques significantly improve solution quality compared to traditional heuristics, with specific examples showing improved results in TSP routing (6659.9 to 6659.431) and more effective resource allocation in complex scheduling scenarios.

Research Overview: Cosine Similarity is a measure of similarity between two non-zero vectors that calculates the cosine of the angle between n-dimensional vectors projected in multi-dimensional space. The mathematical foundation lies in the dot product of vectors divided by the product of their norms, resulting in values ranging from 0 to 1 (or -1 to 1 for semantic variations). Unlike magnitude-based metrics, cosine similarity focuses on document orientation, making it particularly effective for text comparison where document length varies significantly. The research demonstrated its superiority over Euclidean distance and word count methods, especially in scenarios where documents differ in size but share similar thematic content.

Core Applications: The research explored multiple real-world applications including plagiarism detection, information retrieval and search engines (ranking documents by query relevance), recommendation systems (matching user preferences with content characteristics), text classification and clustering (grouping similar documents), sentiment analysis (comparing texts with predefined sentiment vectors), and automatic document summarization (identifying relevant sentences). The study highlighted how cosine similarity's efficiency in processing vectorized representations makes it suitable for large-scale datasets and real-time comparisons, contributing to academic integrity and content originality preservation.

Advanced Techniques for Improving Cosine Similarity: The research investigated numerous advanced techniques to enhance cosine similarity performance: (1) TF-IDF Weighting - Term Frequency-Inverse Document Frequency weighting reduces the impact of common words and emphasizes distinctive terms, significantly improving similarity accuracy; (2) Word Embeddings - Techniques like Word2Vec, GloVe, and FastText create dense vector representations that capture semantic relationships, enabling similarity calculations that understand word meanings beyond exact matches; (3) Document Embeddings - Doc2Vec and Sentence Transformers generate document-level embeddings that preserve semantic context, allowing for more nuanced similarity comparisons; (4) Preprocessing Enhancements - Stemming, lemmatization, and intelligent stop word removal normalize text while preserving meaningful content; (5) Dimensionality Reduction - Principal Component Analysis (PCA) and Singular Value Decomposition (SVD) reduce computational complexity while maintaining similarity relationships; (6) Advanced Similarity Metrics - Soft cosine similarity accounts for semantic relationships between features, while adjusted cosine similarity handles user bias in collaborative filtering scenarios; (7) Context-Aware Embeddings - Transformer-based models like BERT, RoBERTa, and Sentence-BERT generate contextually rich embeddings that understand word polysemy and sentence structure, dramatically improving similarity accuracy; (8) Hybrid Approaches - Combining multiple techniques such as TF-IDF with word embeddings, or ensemble methods that aggregate results from different similarity calculations, provides robust and accurate similarity measures across diverse text types and domains.