At Atheon, I fine-tune Claude 3 Haiku on AWS Bedrock and evaluate responses for groundedness and faithfulness. I also improved retrieval Recall@5 from 0.70 to 0.84 across 300 expert-reviewed queries.
I deploy Dockerized FastAPI and LangGraph workflows on AWS, using validation, human approval gates, and distributed tracing to reduce failed staged executions from 12% to 5%. I also built pipelines for DeFi risk scoring that improved temporal-holdout PR-AUC from 0.48 to 0.61.
At Auraria Campus, I built data pipelines that brought support tickets and device records into BigQuery, reducing weekly data preparation from six hours to 80 minutes. My forecasting models reduced weekly support-demand MAE from 17.4 to 13.8 tickets.
At Xythum Labs, I developed machine-learning experiments and risk-scoring workflows, reducing false-positive alerts by 21% while maintaining 88% recall. In my LLM Reliability Evaluation & Council Chat project, I built Python and Rust evaluation pipelines that improved failure-detection F1 from 0.71 to 0.83.

