At SPAR, I’m conducting concept-injection experiments to investigate whether LLM self-reports reflect internal model states or only the supplied context.
In my independent software engineering work, I’m developing code-defined agent pipelines and evaluations based on recorded engineering decisions. After prompt and RAG approaches plateaued at 65% agreement with my rulings, I expanded the work to fine-tuning a model.
I also designed and built a hybrid RAG system for coding agents, combining graph and vector retrieval with source-version history to assemble related context in a single request.
At Chegg, I replaced an unreliable AI-based student scheduler with a deterministic TypeScript implementation using a min-conflicts algorithm and a schedule-conflict verifier. I also built tools for upgrade testing and support workflows, and adapted a legacy Python application during an Auth0 migration.
