• Identified that large language models struggled with 30-40% lower accuracy on graph reasoning tasks due to suboptimal node presentation order in context windows
• Implemented three distinct graph traversal algorithms (BFS, DFS, custom heuristic-based ordering) to systematically explore 500+ different ordering strategies across benchmark datasets
• Engineered a Python-based evaluation framework using LangChain and GPT-4 to automatically test and measure performance improvements across 1,000+ graph reasoning queries
• Achieved a 28% improvement in task accuracy by optimizing node ordering, demonstrating that context organization significantly impacts LLM reasoning capabilities for structured data

⚙️ Tech Stack:

• Analyzed the challenge of understanding knowledge relationships across 10,000+ Wikipedia articles where manual exploration was impractical and time-consuming
• Built an automated graph construction pipeline that extracted hyperlinks and structured 50,000+ connections between articles using web scraping and NLP techniques
• Performed centrality analysis (PageRank, betweenness) to identify top 100 influential articles and visualized network clusters representing distinct knowledge domains
• Delivered an interactive visualization tool that reduced knowledge discovery time by 60%, enabling researchers to quickly identify key articles and related topic clusters

⚙️ Tech Stack:

• Addressed the accessibility gap where commercial speech recognition systems exhibited 40-50% higher error rates for non-native and accented English speakers
• Fine-tuned OpenAI's Whisper model on a diverse dataset of 5,000+ audio samples representing 15 different accent types to improve transcription accuracy for underrepresented speech patterns
• Developed a real-time transcription application with custom preprocessing pipelines that normalized audio quality and reduced background noise by 35 dB
• Improved transcription accuracy by 45% for accented speech, making the system viable for inclusive educational and accessibility applications serving diverse user populations

⚙️ Tech Stack:

• Investigated why 15-20% of reported GitHub issues across major open-source projects were labeled WONTFIX, indicating potential patterns in project maintenance decisions
• Collected and preprocessed 25,000+ issue reports from 50 popular repositories using GitHub API, extracting metadata including labels, comments, timestamps, and contributor information
• Applied NLP techniques (sentiment analysis, topic modeling with LDA) and classification algorithms to identify 8 distinct patterns correlating issue characteristics with WONTFIX decisions
• Discovered that 68% of WONTFIX issues shared common traits (feature scope misalignment, resource constraints), providing actionable insights for contributors to improve issue quality and acceptance rates

⚙️ Tech Stack: