Dual-agent architecture
Two task-specialized agents with separate prompts, tool sets and isolated ephemeral memory: one for autonomous zero-clarification data entry, the other for cautious analytics querying. Splitting them meant neither had to compromise its behaviour for the other's use case.
Natural language to SQL
The hard part here is backend, not prompting. Dynamic database schema introspection with a five-minute cache feeds query generation, and every generated query passes runtime security validation enforcing multi-tenant data scoping before it executes.
Guardrails
Client-side UUID validation before API mutations, pre-flight duplicate-entity checks, and agent-readable error messages that guide self-correction rather than failing opaquely.
Retrieval
A RAG pipeline over two Pinecone vector indexes with resumable embedding scripts and metadata-size fallback handling for large nested records.
Status
Discontinued before production deployment. The engineering is real; the product never shipped.