Descriptions:
David Park, head of applied AI engineering at Landing AI, presented at AI Dev 26 in San Francisco on the architectural principles behind building reliable, production-grade agentic systems — using a real-world loan origination pipeline as the central example. Rather than a product demo, the talk is a systems-level discussion of what breaks when agentic solutions are scaled to enterprise environments handling regulated, high-stakes decisions.
The architecture Park describes uses a hierarchical multi-agent structure orchestrated by Google ADK, with Claude Haiku handling reasoning and natural language interaction, and Landing AI’s Agentic Document Extraction (ADE) platform parsing unstructured inputs including PDFs, Excel files, and Word documents. A single manager agent holds decision authority — approving, denying, or escalating to human reviewers — while specialized downstream agents apply deterministic business logic after the LLM-driven extraction phase is complete.
Core design principles include separating ‘the brain and the hand’ (a pattern from Anthropic’s engineering blog), wrapping all model calls in typed interfaces with validation and retries, and keeping post-extraction steps fully deterministic to maximize auditability. Park argues this pattern — stochastic LLM cores inside deterministic shells — generalizes across financial services, legal contract review, insurance claims processing, KYC onboarding, and equity research wherever documents are the source of truth for consequential decisions.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Workflow Case Study
