Miro AI provides internal teams with high-leverage building blocks for intelligent features: retrieval-augmented generation, structured agents, evaluation tooling, and guardrails. This portal documents how we build, operate, and evolve those capabilities safely.
Goals
- Enable product teams to ship AI features quickly with strong defaults.
- Centralize safety, logging, and evaluation so quality improves over time.
- Make experimentation cheap while keeping production reliable.
Main Components
- Gateway: normalizes requests to model providers, enforces quotas, and injects guardrails.
- Registry: tracks prompts, versions, datasets, and evaluation runs.
- Orchestrator: agent runtime for tool use, grounding, and post-processing.
- Observability: traces, redactions, cost accounting, and feedback loop.
Mental Model
Treat each AI-powered workflow as a product surface with inputs, policies, runtime dependencies, and measurable outcomes. Changes (prompt edits, model swaps, tool wiring) must be reviewed and evaluated in lower environments before reaching production.
Note
We use model-agnostic interfaces. Upgrading a provider should not require application teams to refactor business logic.