We built the graph engine ourselves.
Nothing off the shelf could hold this.

The honest answer is that we tried. General purpose graph databases store a generic node with a generic property bag, and pay for that generality on every hop. At internet scale, where the interesting answers are four and five hops out, that cost is the difference between a query and a coffee break.

So the engine models internet objects natively. A prefix knows it is a prefix. An autonomous system knows what it announces. A certificate knows every name it covers. The traversal never has to ask what it is holding.

Cypher, properly. A real lexer and parser, a planner that pushes type predicates down and rewrites scans into index lookups, and compiled row iterators that allocate nothing per row. Plans cached in two tiers.

It is not a translation layer over something else. Which matters, because the moment your queries get interesting is the moment a thin wrapper stops keeping up.