SOTA model for high-precision product identification

Company

ShelfSet LLC

Industry

Retail

Country of the Company

USA

Type of Service

Development

The opportunity

  • The client required a high-performance Object Detection model to replace their existing pipeline, which was limited by licensing and accuracy issues.
  • The primary goal was to develop a custom SOTA model capable of identifying products in complex retail environments with high precision, specifically for on-shelf drink identification and planogram monitoring.

Our approach

  • We developed a state-of-the-art model from scratch, training it on diverse, high-complexity datasets to ensure 20% better performance than existing market alternatives.
  • Our work focused on the Data Pipeline, Core Detection Model, and Model Optimization.
  • We delivered the final codebase and repository under an MIT license, optimized for further integration into mobile or web environments.

Value created

  • Our YOLO26s model outperformed ResNet-50 benchmarks for densely packed items.
  • We achieved a 20% boost in accuracy: mAP@50-95 0.59 (vs 0.49), Recall 0.83 (vs 0.55), and Precision 0.90 (vs 0.83). This enables automated “share-of-shelf” analysis and instant planogram tracking with minimal human error.
  • By delivering a scalable detection engine, we’ve built the foundation for the upcoming SKU Classification phase.

Technologies Used

SOTA Object Detection Architectures

PyTorch for training & experimentation

MLflow

TFLite

Albumentations

Albumentations

Model quantization

Company

ShelfSet LLC

Industry

Retail

Country of the Company

USA

Type of Service

Development

Learn about our impact through case studies

Smart app for residential parking sharing

Built a credit-based parking sharing platform that helps residents turn unused parking spaces into a shared community resource, matching available spots with guest vehicles while keeping owners in control.

Real-time AI digital human platform for interactive customer experiences

Building real-time, human-like AI interactions through digital avatars capable of communicating with users and performing practical actions across customer support, sales, and other business scenarios.

AI-powered client engagement and personalization agent

Successfully deployed a secure and compliant AI agent that recommends content, manages schedules, and engages clients in personalized conversations via the fitness-tech platform.

AI agent for real-time price prediction and automated trading insights

Reduced manual intervention in forecasting and reporting workflows with an AI agent capable of handling real-time price predictions, executing analysis, and sending back KPI reports.

AI agent for sterility monitoring in pharmaceutical manufacturing

Enabled fast, accurate detection and escalation of non-compliant procedures, reducing human oversight burden by deploying a real-time sterility monitoring system.

AI assistant for enhanced online learning experience

Digitized and structured all course materials from PDF-documents and videos based on defined business logic, reducing time spent searching by over 70%.