Developing Visual AI Solutions for Online Marketplace
summary
Velebit AI builds end-to-end visual features for marketplaces that run in production not just in papers: a classifier calibrated with temperature scaling and a tree-aware algorithm that returns up to three nodes to trade specificity for accuracy over thousands of leaf categories, a live mobile/video app Babli that skips frames and uses image similarity matching for price suggestions instead of direct regression, an image similarity API with binary vector descriptors and an Elasticsearch plugin, PyTorch → ONNX/TensorRT on Nvidia GPUs in GCP/AWS, and measured gains in listing speed and engagement.
AI-written summary · about
chapters (13)
- 0:00Visual AI for marketplaces
- 2:15AI consultancy agency
- 4:03Key clients and focus
- 4:24Image/text categorization challenges
- 7:57Model confidence calibration
- 10:24Tree-based categorization
- 11:29Real-time video predictions
- 14:57Image similarity for recommendations
- 17:06Robust image descriptors
- 19:36Technical stack overview
- 21:20Other AI projects
- 24:07Business approach and success
- 25:09Q&A
related talks
30mEvolving Search in Online Marketplaces: Practical Applications of AI
31mAI Search for Marketplaces: Experiences & Tech
22mHarnessing image and text data for product matching
16mSolve for AI
33mMachine Learning in Visual Search
30mAI in the Cloud
32mComputer Vision IRL: From Idea to On-Premise Deployment
01h 02mAI on the Edge