Kavach journal
Ideas for systems that make decisions.
Notes on the architecture, protocols and operating principles behind trustworthy AI systems.
Featured writing
From the control plane
Replay is not retry: one word that changes the integrity of the evidence
A retry asks whether work eventually finished. A replay asks whether a governed system can reproduce and defend a historical decision.
Build with clear boundaries
Kavach separates governance state from runtime execution, making policy and evidence easier to reason about.
Protocols are interfaces
The right transport depends on who—or what—needs to consume the capability.
The incident report is not evidence
When an AI decision matters, logs and dashboards are not enough. The system must preserve the evidence required to reconstruct the decision itself.
The hard scaling problem in AI is deciding what to remember
Millions of historical executions do not call for a bigger dropdown. They call for a clear boundary between search projections and authoritative evidence.
MCPO vs Streamable HTTP: choosing the right interface for Kavach
MCP transports and REST adapters solve different integration problems. Here is how to choose between them—and why Kavach can support both.