This repository contains my project submission for the Agentic AI nanodegree program of Udacity.
A multiagent workflow is implemented to process customer requests.
An orchestrator agent analyzes customer requests and coordinates specialized worker agents by calling them sequentially to check inventory stock status, calculate the quote (price) for the requested items, and order restock if necessary.
Workflow diagram:
Customer inquiries enter through the Paper Factory Orchestrator agent, which maps requests to catalog SKUs and coordinates the specialized worker agents by calling them sequentially as tools.
Data flow summary:
- The Orchestrator maps the customer request to up to four catalog SKUs and calls
run_inventory_agent. - Stock Agent checks stock levels per item and places future-dated supplier purchase orders for shortfalls within cash limits.
- Quote Agent builds priced lines using historical context, current cash, and
tool_price_builder; a deterministic patch step corrects LLM output before returning. Sales for immediately available items are recorded at this stage viatool_record_sale. - Fulfillment Agent receives the ready quote lines, calls
tool_financial_snapshot, and composes a customer message that covers confirmed items, backorder ETAs, and delivery commitments — without exposing internal cash figures or margins.
project_starter.py includes run_test_scenarios(), which processes all rows in quote_requests_sample.csv through the four-agent workflow and outputs final balances and customer responses to test_results.csv. Execution depends on the Udacity/Vocareum API key since all agents use pydantic-ai with OpenAI connectivity. (Configure UDACITY_OPENAI_API_KEY in .env then run python project_starter.py.)
- Use a partially nonlinear workflow - this would speed up the processing. For example the quote calculation or restock planning for separate items can be processed simultaneously.
- Handle transactions and cash balance more systematically