Fuseideas

Fuseideas is a Boston advertising agency. I built their website as a headless Next.js and WordPress platform.

As generative AI rapidly became central to marketing and search, I proposed adding a conversational assistant to the site to demonstrate Fuseideas' capabilities in this space. The goal was to allow visitors to explore the agency's services and portfolio through natural language rather than relying solely on traditional navigation.

Editors also needed the ability to manage complex pages without engineering support. WordPress manages content through custom Gutenberg blocks, while the Next.js frontend fetches data via GraphQL and renders components through a typed block registry.

AI Assistant Architecture

The assistant uses a custom RAG knowledge layer backed by Supabase. Agency portfolio content is indexed into the database and retrieved as context before each model response, grounding answers in real work and services rather than generic model output.

To support ongoing maintenance, I built an internal dashboard that allows administrators to index new content and manage the data available to the RAG system.

GPT-4o streams the user-facing response, while a second GPT-4o-mini call runs silently to extract structured signals such as relevant work slugs, capability tags, and suggested follow-ups. These signals drive card rendering and expert routing within the interface.

Conversation history is stored in Supabase and passed into each turn so the assistant maintains session context.

AI Discoverability / AEO

In parallel, I worked on improving the site's AI discoverability. At my digital director's request, I strengthened semantic structure and structured data output to support answer engine optimization (AEO), including improvements to page semantics and JSON-LD designed to make agency content easier for AI systems to interpret and retrieve.