FreshPress AI Assistant (Pressy)
A production-grade AI assistant for a laundry business that quotes live prices, tracks orders in real time, answers FAQs, and escalates to humans — all connected to live Supabase data.
The Problem
Most AI chatbots hardcode pricing and policies into prompts or training data. The moment anything changes, the assistant becomes unreliable — hallucinating prices, inventing policies, and having no idea what is happening with a customer's order.
What I Built
Built a two-layer architecture separating static knowledge from live operational data. A RAG knowledge base handles FAQs, policies, delivery rules, and service information embedded into Supabase pgvector using OpenAI embeddings and n8n orchestration. Live pricing, order status using LAU-XXXXXX order IDs, and conversation history are fetched fresh on every message. Every message flows through: Webhook, Validation, Session History, Live Data, Intent Detection, Order Lookup, Context Builder, GPT/RAG Agent, then Response Delivery. The assistant responds in structured JSON with reply, suggested actions, and escalation flags. The frontend was built in Lovable and connected to a React website with persistent chat sessions, Supabase chat history, dynamic action buttons, and full error handling.
The Result
The assistant quotes live laundry prices, tracks orders in real time, answers operational FAQs, escalates to humans automatically, and remembers conversations across sessions — all without a single hardcoded response. Pricing updates in Supabase reflect immediately with no retraining or prompt edits required.