AI & Graph

Intelligent systems, built on graphs.

AI that reasons over structure, not just text — agentic workflows, GraphRAG on Neo4j, graph-native tokenomics, and graph agent memory. Some of this is shipped and running; some is on the frontier and in development. We label which is which — proof, not promises.

Shipped & running

Agentic workflows

Deterministic multi-agent orchestration — poll → execute → validate → unblock, with a zero-trust validator between every step.

Agent-OS: 12 agents · 755 tasks · 98.7% autonomously verified · $0 infra. MB-agentic governance engine: 1,345 tests, 100% coverage.

GraphRAG on Neo4j

Retrieval that beats flat vector search — hybrid vector + full-text over a knowledge graph, grounded and cited.

Our EcoGraph: 77k chunks + a 9,828-entity knowledge graph of the whole Neo4j corpus. Hybrid retrieval benchmarked +18 pts, recovery doubled vs naive vector.

Graph-native tokenomics

Token economies modelled as graphs and fused with Monte-Carlo simulation — the graph⇄sim bridge that answers structural questions about dynamic output.

Neo4j + GDS + radCAD, applied to real token models (DevGuild $GUILD). See the Portfolio case study.

RAG at the edge

Grounded retrieval running on Cloudflare Workers AI + Vectorize, with MCP tool bridges so agents act on real systems safely and observably.

In production across the studio's edge stack (Kointel, KTHULHU).

On the roadmap

In development

Neo4j agent memory

Graph-backed long-term memory for agents — entities, relationships and episodic recall in Neo4j rather than a flat vector store.

AI-agent GraphRAG

Production agent retrieval over live knowledge graphs — the hybrid retriever we benchmarked, packaged as an agent capability.

Agent-ID & Web3 agent security

Identity, authorisation and security for autonomous agents acting on-chain — a fast-growing surface as agents get wallets.

Graph-native security-audit engine

Smart-contract auditing on a Code Property Graph (AST + CFG + DFG) with a multi-agent roster — Topology → Vulnerability → Exploit → Verification → Reporting — and an EVM-sandbox auto-heal loop. The direction for the KTHULHU engine.

Industry context (attributed, not our metric): an independent study reports agents using GraphRAG are ~80% more truthful than vector-only RAG, answering roughly 2× the questions with fewer hallucinations — “Independent study: GraphRAG makes AI agents 80% more truthful”, Neo4j. Our own measured result is the +18-pt hybrid-retrieval lift above.

The through-line

Systems are graphs. Model them, then prove it.

An LLM over flat text guesses; an LLM over a graph reasons about structure — who connects to whom, what depends on what, which path an attacker or a whale can take. The same discipline runs through everything we build: model the system as a graph, query it, then prove the constraints hold.

+18pts
grounded-answer lift from hybrid GraphRAG over naive vector search — our own benchmark, not a vendor claim.
See graph modeling in a real case study

Building AI that has to be right?

Deterministic agents, GraphRAG grounded in your data, or a graph model of a system you can query. Tell us what you're building.

Book a technical call