I’m sharing Kakveda, an open-source project that explores failure intelligence for LLM and agent-based systems.

Most observability tools help analyze failures after they happen. Kakveda treats failures as first-class entities that can be remembered, matched, and used to warn before repeating known failure modes.

Key ideas: - Global Failure Knowledge Base - Deterministic failure fingerprints - Pre-flight “this failed before” warnings - Failure pattern detection - System health scoring over time

The project runs locally via Docker Compose and does not require a hosted service.

Docs: https://kakveda.com

I’d appreciate feedback on the architecture and ideas.