Retail-Execution Intelligence Platform

Problem

Enterprise FMCG and mobile-financial-services clients needed automated, at-scale verification of point-of-sale material (POSM) placement, planogram compliance, share-of-shelf, and competitor presence across huge outlet networks — work that was previously manual, slow, inconsistent, and impossible to audit at national scale.

Approach

• 01Built a central FastAPI inference service that routes each outlet image to campaign-specific YOLO detection/segmentation models (PJ Strike, Autograph, Euphoria, Frost, Luckies, Derby and more for BAT) rather than one monolithic model.
• 02Scored planogram compliance, blanks, and competitor placement, and added face-recognition and NLP modules for field-agent verification and audio brand-mention checks.
• 03Wrote per-client converters mapping AI class names to each client's product IDs and data schemas, plus cross-model reconciliation for Unilever's overlapping detectors (Display Audit, QPDS, SOS, MTSOS).
• 04Deployed from Hugging Face model repos under Hypercorn with async I/O, structured JSON logging, and HTTP-Basic-secured docs for GPU/container hosting.

Results & Impact

• ~99% detection accuracy across a network cited at 2M+ outlets for enterprise FMCG & MFS clients.— CV-sourced
• BAT_Master flagship sustained 397 commits over 18 months across 7 branches — a mature, continuously-shipped production system.
• Unilever counterpart runs 10+ versioned YOLO models with cross-model reconciliation; parallel live deployments for Malaysia, Square, and Nagad (MFS branding).