Internal reference systems and prototypes that show how we build production AI — architecture, what gets logged, where evals run, and the failure modes we design for. Clearly labelled; not client case studies.
A control layer for agent runs, costs, evals, approvals, and failures.
Ingestion to hybrid retrieval to reranking to cited answers.
A bounded LangGraph agent with explicit state, tools, and approvals.
Repo instructions, skills, MCP, and a GitHub Actions review loop.
A TypeScript MCP server exposing internal tools behind auth.
Routes tasks across OpenAI, Claude, Gemini, and Grok with fallback and cost control.
DVNC Experiments
Standalone concepts where interaction is part of the argument. Each experiment is public, clearly bounded, and separate from client work.
Public concept · Version 3
Choose whether one fictional refund request should be answered, actioned, approved, handed to a person, or refused. The interface shows why those outcomes need different authority and evidence.
Open Authority Relay