I fine-tuned Qwen2.5-Coder-7B-Instruct with QLoRA on GitHub code-review data to generate actionable review comments and identify code that needs no changes.
For that project, I built an evaluation pipeline comparing the base and fine-tuned models, then integrated Claude as an LLM judge. I deployed the system as a FastAPI REST API and CLI for local inference.
On DocsRAG, my personal project, I built a hybrid assistant for technical documentation using dense vector retrieval, keyword search, and the Anthropic API. I added cross-encoder reranking and code-aware chunking, and evaluated retrieval hit rate, faithfulness, and relevance.
As an AI Engineering Intern at Advaith Innovation, I built ML data pipelines and Python/SQL preprocessing workflows for model-ready datasets. I also optimized MySQL queries and backend workflows, reducing pipeline noise by 15%, and created Python and Tableau dashboards for pipeline and model-performance metrics.
