Health-product Compliance & Pharma E-commerce

Two briefs. Real-shaped work.

1. Project 1

   Compliance checker for supplement and ayurveda listing claims

   Build a claim-checking workflow for health supplement and ayurvedic product listings on marketplaces. It reads each listing's title, bullets, description and image text, classifies the product's likely regulatory category, tags every claim by type, flags DMR Act scheduled diseases and other red flags, and writes draft findings with reasons to a review sheet. It never edits or publishes a listing. You then measure it against a gold set you labelled blind, fix what it gets wrong, and hand over an SOP that names the human who signs off. This is a teaching exercise, not legal advice.

   You hand in

   • A working checker (n8n workflow export or a documented Sheets-plus-prompt pipeline) that writes DRAFT findings only
   • A red-flag lexicon of at least 40 terms with Hinglish variants and a reason for each
   • A blind-labelled gold set of at least 40 synthetic listings with your labels and reasons
   • A confusion table with precision and recall before and after one round of fixes, plus an error analysis of every miss
   • A one-page SOP covering scope, escalation, sign-off and what the checker must never do
   • A 3-minute recorded walkthrough for the viva

   Data: Synthetic supplement and ayurveda listings from the Saarthi sandbox (no real brands, no real sellers), plus synthetic label transcripts. No real customer or seller data.. Reviewed by Saarthi against a published rubric, signed off by a mentor.

2. Project 2

   Label-and-claims audit of 20 wellness SKUs

   Audit 20 wellness SKUs from the anonymised Saarthi seller catalogue (frozen dataset version). For each SKU, check that the listing shows every mandatory Legal Metrology declaration and that it matches the label transcript, and review every claim against its category's rules. Use AI to extract and pre-check, then verify every finding yourself. Your counts of missing declarations, mismatches and high-risk claims per SKU are compared automatically with the Saarthi engine's reference output for the same dataset version, and your written findings are graded against the rubric.

   You hand in

   • An audit sheet with one row per SKU per declaration and per claim, showing listing value, label value, status and evidence
   • A per-SKU summary in the required CSV format: sku_id, missing_declarations, mismatches, high_risk_claims, overall_risk
   • A findings report with one finding per issue: SKU, rule, evidence, risk, suggested fix and reviewer decision
   • A log of AI extraction errors you caught and corrected
   • A one-page memo to the (synthetic) seller ranking the five most urgent fixes in plain language

   Data: Anonymised Saarthi seller catalogue, wellness category, frozen dataset version published with the brief: 20 SKUs with listing fields and label transcripts. Seller and brand identities removed.. Reviewed by Saarthi against a published rubric, signed off by a mentor.