A building design consultant ditched a costly Wix site, spending four months developing an AI 'talker' to handle frequently asked questions and client interactions. The motivation stemmed from frustration over paying $40/month for a static brochure that required the owner’s time for basic support tasks. This new system aims to allow the operator to focus on high-level consulting work.

Building the agent involved a heavily modified stack, architected around Netlify’s restrictive 10-second serverless timeout. This constraint forced the developer to segment the logic into three distinct components: the Brain, Hands, and Voice, all running on the Edge. The developer, returning to coding after three decades, heavily relied on AI assistance for the complex build.

Proving its capability, the agent, powered by DeepSeek-R3, successfully defended the business model against a direct challenge from a licensed architect in a public log. A major engineering hurdle involved tuning the LLM over 2.5 months to switch rapidly between a warm client-facing persona and a defensive, caustic tone when provoked.

Practical issues persist, however, including unpredictable behavior from the Web Speech API with non-English inputs and massive liability concerns surrounding potential hallucinations of building codes. The developer opted for an 'Eager RAG' token-burning hack to boost responsiveness, deliberately sacrificing persistent user data for speed, as few visitors return to the site.