Griffiss Institute / Air Force Research Laboratory
AI Research Fellow · May 2026–Present
- Built a reusable Python pipeline to acquire roughly 5,000 research papers from OpenAlex and Semantic Scholar, parse them, and extract structured evidence with LLMs, supporting literature-review automation and institutional trust research.
- Benchmarked Python PDF-parsing libraries against LLM-based extraction on accuracy, segmentation quality, hallucination rate, and cost; the results set the pipeline’s defaults.
- Built reproducible ML and LLM evaluation workflows on SQLite, DuckDB, and PostgreSQL, with Qdrant for vector search, Pydantic schema validation, and resumable batch processing across multi-million-row experiment datasets.
- Configured a two-node NVIDIA DGX Spark cluster for on-premises edge AI inference, benchmarking tensor- and pipeline-parallel vLLM serving of open-weight LLMs across concurrency levels to identify the highest-throughput configurations.
- Set up a standardized development environment on DGX Spark hardware for incoming interns and contractors, automating the provisioning of Open WebUI, Tailscale, VS Code, and coding agents to mirror on-premises servers.
- Designed and analyzed experiments in Python and R, tracking configurations in MLflow and scoring clustering quality, stability, and agreement against reference categories.
- Applied natural language processing, embeddings, clustering, and network analysis to evaluate and refine a computational taxonomy of psychological influence and propaganda techniques affecting institutional trust.
- Developed an agentic-development methodology for the team, defining project intake, agent roles, workflow decision trees, LLM skill files, and quality gates.
