Capstone: AI Workflow Automation Platform

Build a platform that combines deterministic workflows with bounded model decisions for document-driven business processes.

System scope

• versioned workflow definitions and durable execution;
• document ingestion and structured extraction;
• model-based classification behind schemas;
• deterministic routing, policy, and state transitions;
• human queues for ambiguity and high-impact decisions;
• idempotent integrations with two enterprise systems; and
• tenant quotas, audit, observability, and rollback.

Loading diagram...

Required evidence

Implement one process such as warranty claims, invoice exceptions, or access requests. Measure straight-through completion, correction, human review load, cycle time, error severity, and cost.

Compare a rules-only baseline, an unconstrained agent, and the hybrid design. State why each model decision cannot be replaced by a simpler rule.

Failure tests

• duplicate events arrive out of order;
• extraction returns valid syntax with inconsistent values;
• an integration succeeds but its acknowledgement is lost;
• policy changes while work is in progress; and
• a reviewer rejects the model's proposed classification.

Completion criteria

The platform passes when workflow state is recoverable, actions are idempotent, ambiguity reaches the right reviewer, and model output never controls authorization directly.