Visual Question Answering using LLMs
Developed a chart QA system using mPLUG-Owl2 with a 2-stage QLoRA pipeline on over 600K examples, improving accuracy by 18% and reducing hallucinations by 25%. Deployed a production-grade system with REST API backend and Gradio UI for real-time and batch inference.
Personalized Offer Ranking
Created an end-to-end learning-to-rank system using LambdaMART and XGBoostRanker, processing over 500K daily transactions for 10K+ users, improving recommendation relevance by 32% and doubling offer-to-conversion rates.
Advanced Search Engine and Corpus Q&A Tool
Built a search engine using the Rabin-Karp algorithm to find substring patterns across 100K+ documents with sub-100ms query latency, achieving 85% precision and reducing LLM hallucination by 30%.
E++ Expression Evaluator and Compiler
Implemented a custom data type for arbitrary-precision integers and built a compiler converting expressions to stack-machine bytecode, achieving a 40% runtime improvement and 25% memory reduction across 500+ test cases.