SupplySafe is a four-agent autonomous supply chain intelligence system that monitors every supplier daily using live web search, traces any detected disruption through the bill of materials and inventory to the affected customer orders, calculates total revenue at risk, and generates draft communications for supplier and customer contacts — all without human instruction.

The system is not a dashboard that displays supply chain data. It is a pipeline that reasons about supply chain risk and produces actionable intelligence and pre-written communications before the manufacturer asks for any of it.

Four Layers of the System

• Layer 1 — Supplier Intelligence (Agent 1): For each supplier in the database, the agent fires a real web search via Tavily. It retrieves current news, logistics reports, and geopolitical signals. Groq's Llama 3.3 70b then reasons about whether the retrieved content represents a genuine disruption risk for this specific manufacturer — not keyword matching, but contextual judgment. Confirmed risks are written as structured alerts to Supabase.
• Layer 2 — Order Risk Analysis (Agent 2): The agent reads confirmed alerts, queries the inventory table for materials supplied by flagged suppliers, cross-references the bill of materials to identify which products depend on those materials, and traces through open customer orders to find every order currently at risk. No LLM required — this is pure Python logic over structured data.
• Layer 3 — Financial Impact Assessment (Agent 3): The agent aggregates affected order values, calculates total revenue exposure, and writes a structured risk assessment to Supabase with a full cascade summary — from the supplier event to the specific customer orders at risk, in plain English.
• Layer 4 — Response Generation (Agent 4): The agent reads the risk assessment and drafts two categories of communication: supplier status request emails addressed to the specific contact at each flagged supplier, and proactive customer delay notices addressed to the specific contacts on affected orders. All drafts enter the approval queue as pending_review — nothing sends without human sign-off.

Technical Stack

What Makes It Agentic

The distinction that matters: in a workflow automation tool like n8n, the developer writes every rule and every branch in advance. The system executes those rules. If something happens that the developer didn't anticipate, the system doesn't handle it.

In SupplySafe, no rule was written for "Shenzhen port congestion." No rule was written for "semiconductor shortage in South Korea." The Supplier Intelligence Agent reads unstructured news from the live web and decides — through LLM reasoning — whether it represents a risk for this manufacturer. That judgment cannot be reduced to an IF/THEN rule. That is what makes it agentic.