Give a gift of sight to your online commerce

Delight your customers with search capabilities and recommendations that can see and understand fashion, style and decor. Let shoppers find products they’re looking for in a snap -  even the things they can’t name. Convert your lifestyle images into showrooms. Take control of your products’ metadata. The opportunities are limitless with AI-powered image recognition.

Our clients

Retail
Hi-tech
Manufacturing
Finance

How our visual search technology works

Find similar images by similar vectors

Visual search AI models represent or “embed” images with their “fingerprints” - series of numbers called vectors that capture the most essential features of an image. Those fingerprint vectors are created in such a way that similar images have similar fingerprints and are geometrically clustered together. This embedding trick allows the search engine to quickly find similar images by their vector representation.

Full-fledged image recognition pipelines

Our computer vision models perform all tasks required for successful visual search - from object detection, to segmentation, to vector representation and vector search of the target image.

Best tools for the job

Our visual search technology utilizes the most recent advancements in computer vision powered by deep learning algorithms. We custom-select, modify, ensemble, train, and fine-tune state-of-the-art visual models for a particular task.

Collect and organize domain specific data

We train our models on all available data related to the task - both customer-specific and publicly available. We perform data cleaning and labeling using advanced unsupervised and semi-supervised techniques to build large scale datasets to achieve the best search results.

We created our visual search blueprint based on large scale implementations in the public clouds as well as on-premise for Fortune-1000 companies. We focused on open source and cloud native software, and state-of-the-art deep learning model architectures to enable seamless deployment on any public cloud or private infrastructure. We partner with AWS, Google Cloud, and Microsoft Azure cloud providers to ensure the highest efficiency and best practices. Accelerate implementation with Grid Dynamics visual search blueprint
Features
  • Accurate results - up to 98 percent item identification accuracy.
  • Advanced similarity - the AI model takes into account fashion, decor, and artistic style.
  • High throughput - battle-tested architecture handling thousands of parallel searches.
  • Low latency - low latency with optimized vectorizers and fast approximate nearest neighbor search.
  • Highly scalable and robust - share-nothing microservices architecture ensures high scalability and resilience.
  • Integrations - data consumption from message queues, databases, or file dumps and REST APIs ensure seamless integration with the rest of the ecosystem.
Technology stack
  • Infrastructure -AWS, GCP, or Microsoft Azure are supported. On-prem solution is available as well. 
  • Deep learning: A choice ofTensorFlow or Pytorch
  • Vector/ANN index: Milvus or Elasticsearch, as well as embedded implementations
  • Data platform: A choice of Apache Spark, Apache Flink, or Apache Beam are the primary choices along with their cloud wrappers.
  • Feature store: Feast is one option, yet many non-specialized databases and EDW solutions will work
  • Infrastructure -AWS, GCP, or Microsoft Azure are supported. On-prem solution is available as well. 
  • Deep learning: A choice ofTensorFlow or Pytorch
  • Vector/ANN index: Milvus or Elasticsearch, as well as embedded implementations
  • Data platform: A choice of Apache Spark, Apache Flink, or Apache Beam are the primary choices along with their cloud wrappers.
  • Feature store: Feast is one option, yet many non-specialized databases and EDW solutions will work
Workshop

We offer free half-day workshops with our top experts in computer vision and visual search to discuss how to apply those emerging technologies to your business and share industry best practices.

Proof-of-Concept

If you have already identified a specific use case for visual search we can start with a 4–6 week proof-of-concept project to demonstrate the power of modern AI-based visual search and recommendations based on your data and your domain.

Discovery

If you are in the stage of requirements analysis and strategy development, we can start with a 2–3-week discovery phase to identify the correct use cases for visual search, design your solution, and build an implementation roadmap.