I'm currently building a project where a lightweight JavaScript SDK is embedded into client websites, allowing them to modify their pages through a side-panel interface. The next phase is to evolve this panel into a conversational AI experience. The goal is to enable users to input natural language instructions such as “add an interactive slider at the top” or “change all headings to a dark blue serif font,” with the system intelligently interpreting the request, identifying the site’s existing design system, generating the necessary HTML/CSS/JS, and applying the changes to the DOM in real time. The scope of supported actions includes: - Creating new elements (e.g. interactive sliders, image galleries, single and multi-step forms) - Modifying existing elements - Removing elements - Updating styles (colours, typography, layout) - Editing content - Injecting JavaScript behaviours and event handling - Other types of modifications A key requirement is that all generated output must align with the site’s existing design tokens to ensure visual consistency. What I’m looking for: - Backend (PHP 8) - A server-side endpoint that brokers requests to an LLM (e.g. OpenAI, Claude) - Clean, framework-free implementation in vanilla PHP - Structured request/response handling for reliability and extensibility Frontend (JavaScript SDK) - Utilities to extract relevant design context (CSS variables, computed styles, key layout patterns) - Efficient context packaging to minimise token usage - Safe DOM injection of AI-generated output Reliability & Safety - Validation and fallback mechanisms to prevent malformed AI responses from breaking the page - Graceful degradation when outputs are incomplete or invalid Implementation - Clear setup instructions so this can be easily integrated into any site already running the SDK - An MVP flow: user prompt → model request → validated response → live DOM update I already have a working version, but it lacks consistency, doesn’t properly leverage the existing design system, and has some stability issues. I’m looking to refine this into a more robust foundation and continue iterating on it long term. I’m happy to share the current implementation and collaborate on improving the architecture, model strategy, and overall reliability.