MCP SERVER
billing-brain
A compiled knowledge graph served through MCP. It filters signals, connects context, traces evidence, and monitors its own health. Every answer carries its provenance. Every claim can be traced to data — or flagged as ungrounded.
Billing knowledge lives in code, design docs, support tickets, BigQuery tables, and people's heads. No single place to ask "what do we know about dunning?" and get a sourced, confidence-rated answer.
Code reading reveals mechanisms. Tickets reveal customer pain. BigQuery reveals trends. Design docs reveal intent. Nobody synthesizes across all four.
When someone says "dunning stabilized after the crisis," you can't ask: where did that claim come from? Is it still true? What data supports it?
Facts go stale. Metrics stop being measured. Theses go untested. Without active monitoring, the knowledge base degrades invisibly.
A knowledge graph compiled from billing source material — code, architecture analysis, ticket digests, BigQuery observations — and served through an MCP server that any agent can query.
The graph is authored from analysis, not extracted by crawlers. Every node was placed deliberately. The compiler validates structure, checks references, and generates health reports.
Every fact carries a status: grounded (backed by data), asserted (claimed but unverified), or speculative (hypothesis). You always know how much to trust what you're reading.
The server monitors its own health — stale metrics, ungrounded facts, untested theses, knowledge debt. It tells you what's degrading before you ask a question that depends on it.
The graph is a model of what we believe about billing, how strongly we believe it, and why.
Four processing layers, organized like a nervous system. Signals enter at the edge and gain context, reliability metadata, and health annotations as they move inward.
Three-tier search finds relevant nodes. Exact match, taxonomy signal routing, keyword search — each result tells you which tier found it so you can weight accordingly.
Walks edges, assembles subgraphs, finds paths between nodes. Builds the context around a result so you see how pieces connect across domains.
Traces evidence chains from claims back to data. Assesses confidence. Detects contradictions. Every answer carries its own trust score.
Health reports, stale metric tracking, knowledge debt ranking, weekly trend analysis. Surfaces what's degrading so you can fix it before it matters.
Nine tools. Three purposes: orient yourself in the graph, find and explore what you need, then verify what you found.
orient — the dashboard. Graph shape, metrics freshness, health warnings, taxonomy overview. Call this first.
search, node, domain, subgraph — locate nodes, expand neighborhoods, survey territories, connect distant concepts.
evidence, health — trace provenance chains, assess confidence, audit the graph's own reliability.
Plus trends for weekly digest patterns and reload to refresh after graph recompilation. Full tool reference →
The knowledge graph spans fifteen domains — from payment collection mechanics to dunning workflows to customer experience gaps. It captures facts, theses, mechanisms, metrics, and the edges that connect them.
Fifteen territories: billing lifecycle, collection, dunning, credits, usage, access, customer, fraud, honest invoicing, migration, reconciliation, and more. Each is a bounded context with its own nodes and cross-domain connections.
Thirteen kinds organized into layers — data (observations, signals, metrics), knowledge (facts, theses, mechanisms), and governance (decisions, risks, questions). Structure nodes group everything.
Every node carries a status. Measured means observed from data. Grounded means supported by evidence. Asserted means claimed but unverified. The status system is what makes the graph honest.
The graph doesn't know everything. But it knows what it doesn't know.