Co-Founder & Technical Lead · May 2026 - Present
West Palm AI
A managed AI service for mid-sized, process-heavy companies: deploy AI assistants and automations around the systems a company already runs, then host, monitor, and support them for a monthly fee.
Visit Live Site →The Challenge
Companies with 30–500 people and process-heavy operations can see what AI could do for them, but don't have the engineers to build it, run it, and keep it working. West Palm AI is the managed service provider for AI: a free assessment to find the opportunity, then assistants and automations built around the systems a company already uses — hosted, monitored, updated, and supported for a monthly fee. I co-founded it with two partners and lead everything technical, plus outbound.
What I Built
I ran discovery with two prospects, wrote proposals for both, and built a working MVP for each to carry the pitch. For a yacht brokerage: an on-demand listing search where a broker types a boat and a Claude agent loop — using server-side web search and fetch tools, so retrieval runs on the model provider's infrastructure instead of a blockable crawler — finds current listings across brokerage sites, and a structured-output pass normalizes them into cards brokers save and annotate. For a construction EPC firm: a bid-review tool that turns an uploaded bid data room into an intake report (table of contents, missing-document checklist, cross-document discrepancy flags citing document and page) and a contract deviation log scored against the firm's acceptable-terms list, with a FastAPI job worker and an accuracy harness benchmarked against a real past bid.
Key Decisions
In the MVPs, the model decides what to search, which pages to read, and how to extract a listing or a clause; plain code decides everything that must be true every time — authentication, spend caps, which rules are in the ruleset, where citations come from, and job retries. As the sole engineer on westpalmai.com, I decomposed the platform into 107 PR-sized specs and built it with a Claude Code agent loop under hard gates: pre-commit lint, typecheck and secret scanning, an 8-job CI pipeline, a 7-day dependency cooldown, and a lint rule that fails any protected route without an authorization test. I also made the site citable by AI assistants with a generated llms.txt, structured data, and answer-first pricing pages.
Impact
Both discovery engagements produced a proposal and a working MVP. The platform — a Next.js marketing site, a Fastify API, a custom learning platform with signed video playback and verifiable certificates, and PostHog funnels — ships with 303 test files and 12 Playwright suites including accessibility and visual checks. Behind it runs an outbound engine I built: dedicated sending domains, warmed mailboxes on a written ramp, and Apollo sequences mixing email and LinkedIn touches.