TLC

TLC is a manufactured-home finance company, and I am its only engineer. I design, build, and run three production web properties: a borrower site serving around 2,500 visitors a month, a portal for a 197-dealer network, and an internal CRM used daily by an 11-person team. They run on TypeScript and React/Next.js over a managed cloud backend, and I own the CI/CD, the infrastructure, and the ongoing production support behind all of it.

The clearest win was the borrower funnel. When I started, inbound was roughly 20 leads a month. I built the acquisition path end to end, the landing pages, the conversion analytics, and the Google Ads integration, and grew it to around 100 a month, a five times increase. Being the only engineer means there is nobody to hand the hard parts to, so the real challenge is keeping three moving properties coherent without a team to absorb mistakes.

One of those properties is a voice intake agent that answers inbound borrower calls, qualifies leads, and hands structured data to the CRM. I built it on Twilio ConversationRelay and LLM APIs, and I designed it around one load-bearing invariant: the agent never owns state and never controls flow. It only proposes structured data, and a separate deterministic layer validates that proposal and decides what actually gets written. The reasoning is simple. A language model in a call path can drop a call mid-sentence, or fill in a field it should have left blank, and in a system a human later trusts, a hallucinated value can never be allowed to become a record. Keeping the model strictly on the proposing side of that boundary makes every call auditable and keeps the records correct by construction rather than by hope.

Beyond the borrower path, I build the internal tooling that automates campaign and lead-nurture work across the dealer network, and I own the technical direction as the properties grow. This is real client work in production: three properties live, the funnel running, and the intake agent taking calls every day.