ReplyIQ: A Prompt-Orchestrated LLM Architecture for Unified Inbox Automation, Sentiment-Aware Routing, and Lead Scoring

This paper presents ReplyIQ, a Next.js/Supabase customer-engagement platform that uses a single hosted large language model (Gemini 1.5 Flash) behind twelve independent prompt-templated API routes to provide reply drafting, sentiment classification, lead detection and scoring, promise extraction, appointment detection, and follow-up generation over a simulated unified inbox spanning WhatsApp, Instagram, and websitechat channels. The system’s relational schema and prompt orchestration pattern are documented in detail, along with – critically – the fact that channel ingestion is currently simulatedby a template-driven message generator rather than connected to live messaging APIs, a distinction the public-facing description of the system does not make explicit. In place of fabricated production metrics, a small, fully disclosed component-level evaluation is reported, executed directly against the deployed code: a robustness test of the regex-based Markdown-code-fencestripping routine that sits between every LLM call and the application’s JSON parsing layer, including a controlled ablationisolating the stripping step’s contribution, and a structural comparison with a more defensive JSON-repair routine takenfrom an unrelated codebase by the same author. The engineering and architectural trade-offs of using prompt engineering over a single foundation model as the entire “AI layer” of a multi-feature SaaS product are discussed, concrete robustness and security limitations are identified, and the system is positioned within the recent literature on LLM-based sentiment analysis and lead scoring for customer engagement.