I developed a low-code platform for evaluating large language models (LLMs), enabling a team of 20 researchers to contribute to benchmarks including red-teaming, text-to-code, and agentic workflows. I also optimized evaluation speed by 4x, reducing benchmark time from 60 minutes to 15 minutes through dynamic parallelization and threading. I created an automated graph-building algorithm for constructing and executing LLM evaluation workflows. Additionally, I engineered a clean, object-oriented Python backend and implemented a user-friendly interface using Streamlit. I designed a synthetic data generation pipeline to enhance LLM training on ServiceNow modules (NOW JavaScript libraries), generating questions and code snippets with Qwen and Llama. I produced over 20,000 diverse samples for ServiceNow’s Text2Code use case, improving LLM context for accurate responses and code generation.

I designed and implemented “Model Card” in a custom platform to assist companies with MDLC (Model Development Life Cycle) of ML models including data governance using an AWS virtual instance. I also worked on “Analytical Insights” with Streamlit for interactive charts and graphs using plotly, whileintegrating IBM AIF 360 to test out multiple forms of mitigating the model’s bias, displaying the results, and identifying the best one. I utilized Tensorflow to run different models and extract the metrics including a confusion matrix and ROC curve Additionally, I tagged training data for identifying different attributes in model code such as hyperparameters, external dependencies, and datasets