Assignment 3: Intelligent E-commerce Supply Chain Manager

UCLA Extension - Agentic AI Course · Week 3: Planning & Reasoning

Build a hybrid planning agent in n8n that decomposes a high-level supply-chain goal into ordered subgoals, mixes classical algorithms (moving average, EOQ, greedy search) with LLM reasoning at the right seams, and produces an auditable plan.

Start here

➡️ docs/assignment.md — full take-home assignment spec, learning objectives, the 12 TODOs with point values, rubric.

Then, in order:

• docs/planning-primer.md — academic framing (HTN, STRIPS, classical vs LLM, hybrid patterns)
• docs/setup.md — local n8n bring-up, credentials, first run
• docs/submission.md — what to submit and how

For the in-class exercise (separate from the homework): docs/in-class.md — fully built Smart Order Router workflow that auto-imports alongside the homework starter; you trigger it via webhook, observe routing, and modify pieces. ~30 minutes.

Quick bring-up

cp .env.example .env
openssl rand -hex 32          # paste into N8N_ENCRYPTION_KEY in .env
docker compose up -d

Then open http://localhost:5678, create a local owner account, and add an Anthropic API Key Header Auth credential. The starter workflow auto-imports on first boot. Detailed walkthrough in docs/setup.md.

Repo layout

.
├── docs/                     spec, primer, setup, submission
├── workflows/
│   ├── supply-chain-manager-starter.json   the homework (auto-imported)
│   ├── smart-order-router-inclass.json     the in-class exercise (auto-imported)
│   └── topic-planner.json                  warm-up demo
├── custom-nodes/             canonical JS sources for the Code nodes
├── data/                     demo CSVs (sales, inventory, suppliers, shipping)
├── docker-compose.yml        single-container n8n + auto-import + data mount
├── build_workflow.py         regenerates the workflow JSON from custom-nodes/*.js
├── .env.example
└── .gitignore

What you build

Five components inside an HTN-shaped orchestration. The topology is given; you fill in the contents.

Component	Type	What you implement
Master Planner Agent	LLM	Hierarchical decomposition + JSON schema with next_subgoal
Demand Forecasting Engine	Code (JS) + LLM	movingAverage, seasonalIndex, then a context-adjustment prompt
Inventory EOQ Planner	Code (JS) + LLM	eoq formula, detectViolations, then an exception-handling prompt
Supplier Performance Monitor	LLM	A 4-dimension scoring rubric with explicit weights
Logistics Coordinator	LLM + classical fallback Code	Soft-constraint prompt + pickCheapestFeasible greedy planner

12 TODOs total (6 easy / 4 medium / 3 hard) — see docs/assignment.md for the full inventory.

Pinned versions

n8n image	docker.n8n.io/n8nio/n8n:latest (verified against 2.19.5)
LLM	claude-sonnet-4-6
Anthropic API version	2023-06-01

Cost envelope

Target under $0.10 per end-to-end run. Each run hits one Switch branch, so single-branch debugging stays under $0.05.