The API and each worker pool use distinct roles and narrowly scoped data access.
Maya architecture
Explore Maya’s intelligence flow and deployed AWS runtime.AWS runtime
Browser and Slack requests cross CloudFront, AWS WAF and the Application Load Balancer, use Cognito identity, and reach Maya's tenant-scoped API on ECS. The API records work in PostgreSQL and sends tasks through isolated SQS lanes to dedicated workers. Workers use Maya's LLM gateway and connector broker to reach approved AI providers and business systems. Secrets Manager, IAM, CloudWatch, KMS, and encrypted storage protect and observe the complete path. Application tasks use public subnets with restricted ingress; databases use isolated subnets.
Tasks, evidence, approvals, Signals and audit history remain outside the model.
Maya owns tenant access, tool policy, budgets, approvals and the final record.
How Maya continuously understands your business
Maya reduces routine event volume before expensive reasoning, investigates important questions in parallel, and verifies outcomes after action.
Business events are normalized, correlated, and deduplicated into candidate Signals. A cost-aware evaluator selects a small number for investigation. An orchestrator runs bounded parallel investigations across approved systems, merges cited evidence, and uses a frontier model only when deeper synthesis is justified. Proposed actions pass through policy and human approval, then Maya verifies the result and records resolution.
DESIGN LAYERS
Clear responsibility at every boundary
The model is one participant in a governed system. Identity, authorization, execution, evidence, and approvals remain application responsibilities.Request path
Browser and Slack requests cross CloudFront, AWS WAF and the Application Load Balancer before reaching tenant-scoped ECS services.
Durable execution
The API dispatches work into isolated SQS lanes consumed by independently scalable interactive, scheduled, and maintenance workers.
LLM Gateway
A Maya-owned policy layer selects approved models, enforces budgets, records usage and latency, and applies controlled fallback.
Data boundaries
Control data, connector reporting data, artifacts, and credentials remain in purpose-specific encrypted stores with workload-scoped access.
EXPLORE MAYA