What's New
A living history of features, fixes, and improvements shipped to the Retail AI platform.
Team Management & Organization Roles
Organizations can now invite team members with granular roles: Owner, Admin, Member, and Viewer. Includes member management UI, invitation flow, and role-based dashboard access control.
- Invite users by email to your organization
- Four permission levels with scoped access
- Pending invitation management
- Organization switcher with role indicator
Advanced Billing & Credit Top-Up
New top-up system with PayPal and Binance Pay (USDT) integration. Real-time credit balance tracking, invoice history, and automatic license generation after successful payment.
- One-click credit top-up via PayPal
- Binance Pay USDT support
- Downloadable invoice PDFs
- Credit cost breakdown by resolution
Hotfix: Inference Queue Deadlock
Fixed a rare race condition in Celery workers that could cause inference jobs to hang indefinitely under high load. Added automatic retry logic and improved worker health checks.
- Resolved Celery task acknowledgment race condition
- Added 3-attempt automatic retry with exponential backoff
- Worker liveness probe now checks GPU memory state
YOLO11 Model Update & TensorRT Optimization
Upgraded inference engine to YOLO11x with custom SKU-trained weights. TensorRT engine files now pre-built for common batch sizes, reducing first-inference latency by 60%.
- 97.4% detection accuracy on SKU110K benchmark
- Sub-500ms inference on standard 1024px images
- Dynamic batch size optimization
- ONNX fallback for unsupported GPU architectures
Dashboard Analytics & Usage Charts
New analytics section with daily usage charts, credit consumption trends, and API request breakdowns. All data is aggregated server-side for fast loading.
- Interactive Recharts usage graphs
- Daily / weekly / monthly aggregation
- Export usage data as CSV
- Per-API-key usage tracking
Public API v1 & SDK Release
Stable release of our REST API v1 with comprehensive OpenAPI documentation, Swagger UI, and official Python SDK. Includes async inference endpoints, webhook support, and presigned S3 uploads.
- Async inference with webhook callbacks
- Presigned S3 PUT upload URLs
- OpenAPI 3.0 spec with Swagger & ReDoc
- Python SDK with async/await support
Custom Model Training Pipeline
Enterprise customers can now request custom model training on their proprietary shelf images. Includes data augmentation, hyperparameter search, and private model deployment.
- Upload training datasets via secure S3 bucket
- Automated data augmentation pipeline
- Hyperparameter optimization with Optuna
- Private model endpoints with org isolation
Bounding Box Visualization Improvements
Client-side canvas renderer now supports category color coding, confidence score labels, and zoom/pan interactions. Results JSON schema updated with per-detection metadata.
- Color-coded bounding boxes by SKU category
- Confidence label overlay toggle
- Canvas zoom and pan gestures
- Improved JSON result schema
Multi-Resolution Inference Support
Inference jobs now support 512px, 1024px, and 2048px input resolutions. Credit costs scale dynamically based on resolution and plan tier.
- Resolution selector in dashboard playground
- Dynamic credit cost per resolution
- Upscale preprocessing for low-res inputs
- Plan-specific resolution limits
Enhanced Image Upload & Storage
Replaced direct S3 uploads with presigned PUT URLs to eliminate cross-platform upload errors. Added support for WebP and HEIC formats with automatic conversion.
- Presigned PUT URL upload flow
- WebP and HEIC format support
- Client-side image validation
- Automatic format conversion to JPEG
Never Miss a Release
Follow our changelog for the latest features, improvements, and fixes.