Ramy — AI Agent System
Architecture at a glance
Cost routing pipeline
Agent graph (simplified)
RAG: Qdrant · Orchestration: LangGraph
The Challenge
Neel Networks needed an AI-powered sales assistant to engage website visitors 24/7, qualify leads automatically, handle objections intelligently, and reduce dependency on the human sales team — while keeping AI API costs under strict control.
The Solution
Built 'Ramy' — a production AI agent with 8 specialized sub-agents orchestrated through LangGraph. The system uses a 3-tier cost optimization pipeline: Rule Engine (free pattern matching) → Claude Haiku (cheap, for simple queries) → Claude Sonnet (quality, for complex conversations). Each visitor conversation is classified across 13 intent categories, scored as HOT/WARM/COLD, and automatically triggers email notifications for qualified leads.
The architecture includes a knowledge base of 196 vector chunks in Qdrant, covering 49 portfolio projects across 17 industries, 51 crawled web pages, 40 FAQs, and 15 service descriptions. The system tracks conversation phases (Discovery → Exploration → Interest → Evaluation → Conversion) and adjusts agent behavior accordingly.
Two dashboards were built: a Client Dashboard for the sales team (leads management, conversation replay, analytics) and a Developer Panel for prompt engineering (prompt editor with version history, cost dashboard, knowledge base browser).
Results
Tech Stack
Try before you book
Talk to the same Ramy system described in this case study: scripted demo, then optional live mode.
Open live AI demo (Ramy) →Related reading
How this kind of system is designed and cost-optimized in practice—written for technical founders and engineering leads.
Read the companion article on the blog →Want something similar?
Let's talk about your project. I'll scope it out and give you a straight answer on what's possible.