Graph AI LLC
We build scalable graph analytics: parallel graph and Markov-chain solvers, spatio-temporal and network analysis, link prediction, and interactive views that make the structure of large connected systems readable.
Many problems are really about relationships: who connects to whom, how something flows through a network, and where it settles. We turn those relationships into graphs and solve them at scale.
Deterministic, parallel solvers for ranking, stationary distributions, shortest paths and centrality on graphs with millions of edges, with bit-identical results on any core count.
Geospatial and time-aware graph analytics: movement, routing and multimodal transport networks, and how they change over time.
Knowledge-graph construction, link prediction and community detection to surface hidden connections across datasets, with explainable outputs.
Graph structure turned into views people can reason about: landscapes, sections, paths and local graph close-ups, always next to the exact numbers.
C++ and Python engines with full test suites, pre-registered evaluations and honest reporting of what works and what does not.
APIs and data pipelines (Parquet, DuckDB, graph databases) so the analytics run where your data already lives.
Three engines built on one idea: data is a graph, and the answers are in its connections.
Markov transition-probability solver over money flows
Ranks 10,000 US stocks and ETFs by the stationary probability of a Markov chain over estimated stock-to-stock money flows, with the Fluxscape landscape, paths and transition-graph views.
Vendor-neutral knowledge graph framework
One browser UI and one HTTP API over multiple graph engines: build, query, explain and visualize a graph through the same workflow on any configured engine, instead of per-vendor consoles and one-off scripts.
kNN-based document entity and relationship extraction engine
A language model asked to build a knowledge graph asserts only what a sentence states, so related entities can end up disconnected. xVector adds a purely additive second pass: documents are split into paragraph chunks and embedded once, a top-k nearest-neighbour query yields chunk pairs, and Shepard inverse-distance weighting scores the hidden ties between the entities they mention. The pass is order-independent and needs no recomputation as the corpus grows, and it is engine-neutral: the same weighted edges go into FalkorDB, Kinetica, ArangoDB or Neo4j.
Paper: arXiv:2609.00387Where does the market's money settle? MarketRank treats stocks as the states of a Markov chain whose transition probabilities are the estimated stock-to-stock money flows, and ranks each stock by its long-run (stationary) probability. PageRank is the best-known instance of the same idea, with links instead of dollars.