The author built a RAG system for financial Q&A using SEC filings and the FinanceBench benchmark. They uncovered a significant discrepancy between LLM-as-judge evaluations and actual performance, leading to lessons on calibrating LLMs for assessment.
The author developed an AI-powered autopilot system to generate evidence-based health content in three languages for their platform, Longevity AI. This initiative aims to scale content production beyond manual limitations while ensuring quality, ethical compliance, and adherence to regulations.
Removing common opening pleasantries from a GPT-4o-mini cold email prompt increased reply rates eightfold, from 2-3% to 8-15%. The author found that AI-generated pleasantries are easily identified by recipients and flagged by spam filters, making direct communication more effective.