M.Video-Eldorado: computer vision across 1,080 stores lifted conversion by 23%
Network-wide conversion rose from 14% to 17.2% in the first year. Top factor: consultant schedule optimization. Before: 3 people in the washing-machine zone scheduled 10:00-18:00. Now: 5 in peak hours (Wed 17-20, Sat 13-18) and 1 in dips (Tue 11-13). Customer waiting time fell 47% — visitors who don't get a consultant in time leave without buying in 73% of cases. Real-time incident detection revealed three things: 14% of thefts are internal (staff), unexpected scale; 31% of "lost" expensive items (left unattended) were cases where the consultant got absorbed with one client and forgot another; checkout conflicts — 22 cases/day on average, of which 4 escalating (previously only escalating ones were tracked). Key challenge: privacy compliance. The team worked with Roskomnadzor for 7 months on approval: face blurring on edge before frames leave anywhere, no direct identification, anonymous session IDs reset every 4 hours, all "events" auto-deleted after 30 days. All documented in a standard now used as reference for the industry.
Contexto
M.Video-Eldorado is Russia's largest electronics retailer: 1,080 stores in 366 cities, 19,700 employees, 2024 turnover 590 billion rubles. Average store: 1,600 sq.m., average footfall: 280 unique visitors per day. Through 2023, visitor behavior data was collected manually: mystery shoppers quarterly, store director reports, aggregated POS stats. This gave a sparse, delayed, often distorted picture.
Problema
The main pain in electronics retail: conversion. A visitor came in, looked at washing machines, left. Why did they leave? Couldn't find a consultant? Didn't like the price? Too few models on display? Competitor next door? Through 2023, answers were built on guesses. Network-wide conversion: 14% (bought of those who entered). World-class competitors: 22-25%. Each conversion point = ~6 billion rubles revenue per year.
Second pain: incident detection. Theft (internal and external), unattended expensive items, potential customer conflicts, checkout queue overflow. Previously — post-fact from security recordings, a week after the event.
Third: workforce optimization. Store managers didn't know which hours which department needed more consultants. Schedules were built "like usual". At peak hours, the washing zone stood without a consultant; the audio zone had two "free".
Solução
M.Video-Eldorado deployed computer vision across 1,080 stores. Each store: 12-24 cameras (area-dependent), 4K, 30 fps. Processing split between edge (NVIDIA Jetson Orin Nano in each store, 67 TOPS) and central data center. Critical because data flow from all 1,080 stores (200K frames per day each) can't be streamed to cloud.
Three models run on edge: YOLOv8 (retail-tuned) detects people and assigns anonymous session IDs; a ReID model tracks visitors across the store without identification (face blurring is mandatory per 152-FZ); pose estimation classifies "behavior" (examines product, looks for consultant, in a hurry, conflict).
Only aggregated "events" go to central DC (3 KB per event, ~80 events per store per hour): "visitor X spent 4 minutes at washing-machine zone, no consultant interaction, left", or "visitor Y left an item worth 87K rubles unattended for 11 minutes". All anonymized, no identity attached.
Four analytics layers run on the data: traffic heatmap (where people linger), conversion funnel (entered → approached product → picked up → called consultant → bought), incident detection (for security), workforce demand (for HR). The store manager sees a real-time dashboard.
Resultado
Network-wide conversion rose from 14% to 17.2% in the first year. Top factor: consultant schedule optimization. Before: 3 people in the washing-machine zone scheduled 10:00-18:00. Now: 5 in peak hours (Wed 17-20, Sat 13-18) and 1 in dips (Tue 11-13). Customer waiting time fell 47% — visitors who don't get a consultant in time leave without buying in 73% of cases.
Real-time incident detection revealed three things: 14% of thefts are internal (staff), unexpected scale; 31% of "lost" expensive items (left unattended) were cases where the consultant got absorbed with one client and forgot another; checkout conflicts — 22 cases/day on average, of which 4 escalating (previously only escalating ones were tracked).
Key challenge: privacy compliance. The team worked with Roskomnadzor for 7 months on approval: face blurring on edge before frames leave anywhere, no direct identification, anonymous session IDs reset every 4 hours, all "events" auto-deleted after 30 days. All documented in a standard now used as reference for the industry.
Lições aprendidas
- Edge processing is mandatory. 1,080 stores × 200K frames — impossible in cloud. Edge sends only aggregated events.
- Anonymous tracking without identification is the only legal path. Face blurring on edge, not after.
- Workforce optimization is the biggest ROI. Demand-driven schedule = +3pp conversion.
- Privacy compliance matters more than tech. 7 months with regulator > ML accuracy.
- Internal theft is underestimated. 14% of thefts are staff. Changes security process design.