CH
Back to projects
Active
Product DiscoveryData EngineeringStakeholder CommunicationKPI Definition

Retail Analytics Platform

Enterprise shelf intelligence for AI-powered retail monitoring

$2M+annual recoverable revenue identified
Product Screenshots
Shelf Intelligence Dashboard — KPI cards showing actionable OOS, completion rate, on-shelf availability, and estimated lost/recovered sales with department performance table

KPI overview with actionable OOS, completion rate, lost sales, and department-level performance breakdown

Financial impact by store showing products affected, OOS duration, estimated lost and recovered sales per location

Financial impact by store with recovery tracking and response distribution

Architecture
DATA PIPELINE FLOWSOURCESCV Scan PortalVendor API3x dailyscrapeETL PIPELINEPlaywright scraperNormalize + mapYAML configloadDATA STORESQLite + summariesCanonical modelmulti-customerserveOUTPUTSStore dashboardsExecutive decksROI calculatorFlask + Chart.jsn8n ORCHESTRATION — 16 WORKFLOWSScheduling · Error handling · Slack alerts · Deck generation · Health monitoring$2M+/yrrecoverable revenue identified0 hrs manual reportingend-to-end automationZero manual opsfully automated pipeline
Problem

Retail teams had no visibility into what their AI-powered shelf monitoring was actually doing for them. Scan data existed in vendor portals but wasn’t actionable—no lost sales calculations, no trend analysis, no way to measure ROI. Store managers spent 5.5 minutes per section on manual shelf sequencing with no data to prioritize.

Discovery

Conducted time & motion studies across pilot stores (10 sections, 2 stores) to establish baseline efficiency metrics. Researched industry benchmarks (Gruen & Corsten OOS studies) for lost sales methodology. Interviewed store managers and operations teams to understand pain points. Analyzed raw scan data to identify gaps between what was collected and what was needed for decision-making.

Solution

Built a shelf intelligence dashboard with drill-down navigation (store → category → item → daily history). Designed a canonical data model abstracting customer-specific variations via YAML configs, reducing new customer onboarding from weeks to configuration changes. Automated the entire data pipeline: 3x daily web scraping via Playwright, data normalization, summary table rebuilds, and weekly executive deck generation via python-pptx.

Decisions & Tradeoffs
  • Chose YAML-based config over hardcoded customer logic — new customer onboarding went from weeks of development to configuration changes
  • Built lost sales methodology on Gruen & Corsten academic research rather than vendor claims — gave credibility in executive presentations
  • Automated ETL instead of manual exports — eliminated human error and enabled 3x daily refresh cadence
Outcomes
  • Built analytics platform identifying $2M+ annually in recoverable revenue across pilot locations
  • Platform adopted as the standard reporting solution across all pilot customer accounts
  • Eliminated all manual reporting — automated 3x daily ETL, weekly executive decks, and trend analysis
  • Multi-customer expansion from convenience retail to grocery in weeks, not months
  • Fully automated pipeline: 3x daily ETL, 16 n8n workflows, zero manual operations
Built with
PythonSQLFlaskChart.jsPlaywrightn8n