Content moderation and safety, built from scratch. Runs fully offline. No cloud, no data exposure.
Released August 2025. Built on a custom transformer architecture with CoreML support for native on-device inference. Powers content moderation inside CosmicChat, detecting harmful prompts before they reach the model.
View on GitHub →Identifies unsafe inputs before they reach any downstream model.
Custom attention layers built from scratch for classification tasks.
Runs natively on Apple Silicon without any external runtime.
Ships inside CosmicChat as the first line of content safety.
A ground-up rebuild with bidirectional attention, broader content coverage, and a new API layer for businesses and developers.
Where EvoMind 2.0 was built for a single use case, 3.0 is designed to be a platform. Clients define what to block. EvoMind enforces it, at speed, entirely offline.
REST API for integrating EvoMind into any platform or pipeline.
Clients configure exactly what content categories to block or flag.
Full context understanding across the entire input for higher accuracy.
Hate speech, NSFW, PII, prompt injection, spam, and more.
Interested in early access or integrating EvoMind 3.0 into your platform?
Get in Touch