Building autonomous agents and shipping ML systems at the intersection of applied research and production engineering. Currently at NVIDIA; concurrently a (soon to be) published ML researcher at UF. I also make tech content — find me at @joestechtalks.

SELECTED WORK

ExperienceCurator AI   private

A "career memory" system built on agent-orchestrated RAG. Ingests documents and codebases, normalizes content, generates embeddings, and serves citation-grounded evidence through a FastAPI/pgvector REST API. Multi-agent resume tailoring and interview prep cut customization time by 30%.

_{Python · FastAPI · Google ADK · pgvector · Docker · React / TypeScript}

EvoChess   → repo

Dagster MLOps pipeline that ETLs Lichess game data into S3, trains a CNN on 9M+ games rated 2000+ elo with PyTorch, tracks experiments in MLflow, and deploys the final model as a containerized REST API on Cloud Run.

_{Python · PyTorch · Dagster · MLflow · AWS S3 · Docker · GCP}

Auditory Masking Risk Prediction   private · published

End-to-end audio ML pipeline built for UF AI Scholars — published in the UF Journal of Undergraduate Research. Synthesized 56,700 warning-signal/construction-noise mixtures, derived mel-spectrograms and acoustic features, and trained classifiers to predict ISO 7731/9533 audibility compliance.

_{Python · PyTorch · UrbanSound8K · DEMAND · mel-spectrograms}

young-marketers-inbox   private · agentic build

Real-time iMessage inbox for influencer campaign management, built entirely with AI agents as a proof of what you can ship with agentic development. BlueBubbles turns macOS Messages into a REST API; Socket.IO events flow through Next.js via SSE; SQLite owns campaign metadata.

_{Next.js · TypeScript · SQLite · BlueBubbles · Socket.IO · SSE}

ON RESEARCH

AI Scholars researcher at the M.E. Rinker Sr. School of Construction Management, University of Florida. Published work on machine learning for predicting whether safety warning signals are audible over construction-site background noise — ISO 7731/9533 compliance at scale. Prior work includes CTGAN synthetic data generation for heat-stress risk modeling.