IRIS
AI/Computer Vision • December 7, 2025
Vision LLMs hallucinate. IRIS forces them to verify. Uses SAM 3 segmentation to ground Claude's visual reasoning with exact counts, coordinates, and proof masks.
Python
TypeScript
Next.js
FastAPI
Claude
SAM 3
Computer Vision
Key Features
- •Vision LLMs hallucinate counts and positions. IRIS forces Claude to call SAM3 segmentation tools and return exact counts with coordinates
- •Spatial reasoning via mask IoU calculations gives precise overlap percentages instead of vague approximations
- •Frame-by-frame video tracking with SAM3VideoModel outputs temporal object count data across the entire clip
- •Every answer includes visual proof with colored masks, bounding boxes, and confidence scores