Miro AI Architecture

The platform uses a layered architecture to separate concerns and keep the system evolvable as model providers change.

Logical Architecture

• Ingress: API Gateway authenticates, rate-limits, and performs schema validation for requests.
• Safety: prompt templates, redactions, safety classifiers, and policy enforcement are applied uniformly.
• Runtime: orchestrates model calls, tool use, and retrieval; logs structured traces with costs and latencies.
• Registry: canonical store for prompts, datasets, eval results, and approvals.

Request Flow

1. Client authenticates and submits a typed request to the Gateway.
2. Policies and guardrails are applied; requests may be blocked or modified.
3. Runtime executes the plan (RAG, tools, model inference), producing a result + trace.
4. Observability pipeline ships metrics, traces, and feedback for analysis.

Design Principles

• Prefer explicit configuration and versioning over behind-the-scenes magic.
• Minimize coupling to specific model providers.
• Make safety features opt-out only via review.