AI Merchandising & Trend Analytics

Two briefs. Real-shaped work.

1. Project 1

   Trend report from public bestseller data → assortment plan with OTB and sell-through math

   A fictional Indian ethnic-wear D2C brand selling at ₹800–1,500 is planning an 8-week drop. Collect at least four weekly snapshots of publicly visible bestseller rankings for one category on two marketplaces, tag them with an AI assistant you have audited, and write a trend report with three evidenced calls. Turn those calls into a line plan with core, fashion and test roles, a good-better-best price ladder and a net size curve, then build a three-month open-to-buy and set the sell-through and weeks-of-cover rules the buyer will use in-season. Present it as if to a sceptical buyer.

   You hand in

   • Snapshot log of at least 240 dated rows with source, rank, title, price and URL, collected by hand from public pages
   • Tagging audit: 30-row hand check with accuracy reported, and the final tagging prompt
   • Trend report (max 4 pages): scope, three calls with evidence and rule-based confidence, a 'not calling' section, risks
   • Line plan with role and tier shares and a per-style size split that totals exactly
   • Three-month OTB planner that reproduces the course's worked example, plus −15%/0%/+10% scenarios
   • In-season rules sheet: sell-through and weeks-of-cover thresholds with the reasoning for each
   • A 5-minute recorded walkthrough answering three sceptical-buyer questions

   Data: Public bestseller and rank snapshots the student records by hand (no scraping, no login, no product photos), plus a synthetic sales and stock plan for a fictional brand generated in the course workbook.. Reviewed by Saarthi against a published rubric, signed off by a mentor.

2. Project 2

   Size-and-fit return-reduction analysis on anonymised return reasons

   Using an anonymised return-reason export from a Saarthi seller account, find which styles and sizes come back for size and fit reasons and recommend fixes a brand can make this month. Classify free-text reasons with an LLM against a fixed taxonomy, verify the classifier on a blind hand-labelled sample, compute return rates by SKU and size with a minimum-count rule, and write a ranked fix plan with a measurement plan. Your return-rate figures are checked against Saarthi's engine output for the same file, so totals must reconcile to the unit.

   You hand in

   • Classified return file with the final prompt and taxonomy definitions
   • Blind-labelled validation sample of at least 75 rows with accuracy and a confusion table of the main disagreements
   • Return-rate table by SKU and size (units sold, units returned, rate) that reconciles to the export's totals
   • Top five problem cells, each with five quoted raw reasons as evidence
   • Fix plan: size-chart changes with garment measurements, fit notes, image or model-info changes, and grading feedback, ranked by effort and impact
   • A 4-week measurement plan naming the metric, baseline and target check date

   Built on Saarthi’s engines. Your key numbers are checked against Saarthi's own returns logic, run on the same anonymised data.

   Data: Anonymised Saarthi seller return-reason export (SKU, size, reason code, reason text, quantity) with matching units sold by SKU and size. Customer, order and seller identifiers are removed; served from the Saarthi sandbox.. Reviewed by Saarthi against a published rubric, signed off by a mentor.