ML & Data Science with AI copilots

Two briefs. Real-shaped work.

1. Project 1

   Order-level RTO-risk model with a leakage audit

   Build a model that predicts, at the moment an order is placed, whether it will end in RTO (returned to origin undelivered), using only information available then: payment mode, pincode prefix, price, SKU, time and customer history. Prove the model is leak-free with a written audit, show it beats honest baselines, and turn it into a tiered policy a seller could run, with a fairness check and a model card.

   You hand in

   • A reproducible notebook or repository with a time-based split and a leak-free pipeline
   • A leakage audit: feature timing table, the leaks found in a naive version and how each was detected
   • A comparison against two baselines and logistic regression, with precision, recall, PR-AUC, calibration and expected cost
   • A tiered policy with cost-based thresholds and impact shown as a range under stated assumptions
   • A regional fairness check and a drift simulation on the shifted month
   • MODEL_CARD.md and a one-page seller brief, defended in a live viva

   Data: Anonymised seller data pack v1: orders.csv, shipments.csv, customers_anon.csv, pin3_region_lookup.csv, plus synthetic orders_shifted.csv for drift testing. Reviewed by Saarthi against a published rubric, signed off by a mentor.

2. Project 2

   SKU demand forecast that beats an LLM, plus a reorder plan

   Forecast weekly unit demand per SKU for the next four weeks and turn it into a reorder plan for a two-week supplier lead time. Your forecast must beat both classical baselines and a large language model's forecast made from the same history, in a rolling backtest, overall and for each SKU segment you define. Handle stockouts, sale events and intermittent sellers explicitly.

   You hand in

   • Data preparation that flags stockout weeks and sale events, with the reasons for each choice
   • A rolling-origin backtest over at least eight cut-offs for seasonal naive, moving average, the LLM forecast and your model, reporting WAPE, MASE and bias by segment
   • The saved LLM prompts and responses, with the identical prompt used for every SKU and cut-off
   • A reorder plan with safety stock tied to backtest error and a stated service level
   • A two-page memo explaining where your model wins, where it does not, and why
   • Peer review of another student's backtest, and a live defence

   Data: Anonymised seller data pack v1: daily_sales.csv (with in_stock flag), events.csv, sku_master.csv, inventory_snapshot.csv. Reviewed by Saarthi against a published rubric, signed off by a mentor.