What I've actually built with it.

Pest OS: a live AI product

Pest OS uses the Claude API and Retell to answer real phone calls for pest and wildlife operators. 97 calls handled in our first customer's first month, six jobs booked without a human touching the phone. The hard part wasn't getting the AI to talk. It was building the spam screening, the escalation logic, the custom routing that optimizes appointment times by location and job type, and figuring out what to do when it gets a call wrong. Full details on the Entrepreneurship page.

GHC Partners: outreach pipeline and the hallucination I caught

I built an outreach system for GHC Partners using the Claude API to draft personalized LP emails from parsed family office PDFs. Every email required partner approval before it went out.

Partway through the build, the drafts looked great. Clean, confident, personalized. Then I checked whether the personalization was actually correct. It wasn't. The model had been inventing investment focus areas that sounded right but weren't in the source documents. I scrapped the automated lines, went through 87 companies by hand against their source PDFs, and ended up with 28 verified focus areas, 59 left blank, and a do-not-contact list flagged in the CRM.

Research and analysis

I used Claude Code to write and debug the Python that processed 23,470 SEC filings for my finance paper. Claude helped draft visualizations for the Callaway deck, which discloses AI usage in the appendix. I use Claude Code daily for building workflows, writing scripts, and working through problems faster than I could alone. The analysis, the strategy, and the writing are mine.