I built the Instamart Customer Intelligence & Support Copilot as an IIT Patna AIML capstone, a four‑part Python 3.11+ pipeline on a synthetic grocery‑delivery dataset, with all model calls routed through the provided mock_llm.py and no external APIs.
Customers (data/clean/customers.csv, 1,200 rows) were cleaned with zero duplicate customer_ids.
Average catalogue price AVG(price_inr)=289.33 INR for the 40 product rows.
Orders (data/clean/orders.csv, 12,309 rows) remain after deduplication.
112 unparseable/corrupted order amounts ('N/A' and 'error') were repaired using the catalog price formula.
Delivery‑time medians: Dairy: 25.0 minutes, Electronics: 25.0 minutes, Grocery: 25.0 minutes, Personal Care: 25.5 minutes, Snacks: 25.0 minutes.
Grand Total INR 2,998,668.80 for the city‑by‑category revenue matrix (Q2).
Correlation between order_amount and discount_pct is r = ‑0.0951 (very weak negative).
The SQLite database data/instamart.db enforces foreign keys and includes four performance indexes on the orders table: idx_orders_customer_id, idx_orders_product_id, idx_orders_order_ts, and idx_orders_status.