I engineered features, automated tests, and code improvements across GitHub repositories as a freelance Generative AI Engineer, using multi-turn LLM-assisted workflows in Linux and tmux. I also evaluated agentic coding workflows by validating tool calls, execution, and generated code.
At Turing, I optimized prompts by evaluating 500+ prompt-response pairs for quantitative reasoning tasks involving chart images from research papers. I designed and tested Chain-of-Thought prompting strategies for complex, multi-step problems.
I built the GeoAI Research Assistant, a hybrid RAG system over arXiv papers that combines SQLite FTS5 search with embeddings and Reciprocal Rank Fusion. I evaluated it with RAGAS and deployed the containerized application on Google Cloud Run, with PostgreSQL request logging and Grafana monitoring.
For my land-use classification study, I fine-tuned and benchmarked 1D-CNN models on Sentinel-2 imagery, achieving a 90% F1-score with ResNeXt. I presented the findings as a conference paper at ICGCE 2024, and separately implemented transformer and ViT architectures from scratch in NumPy and PyTorch.

