GeometriX capabilities trace to published methods and reproducible benchmarks. The research is organized by theme, and the themes are the ones the products are built on.
Measuring LLM-as-Judge reliability against graph-verified ground truth in financial documents (2026).
A benchmark for structured information retrieval from financial documents using graph-verifiable questions (2026) — ships in the platform as a runnable benchmark.
Gradient-based space-filling designs with an application to systematic evaluation of LLMs on financial-regulatory documents (2026).
The adapter-mesh architecture that federates heterogeneous knowledge backends (2026).
Inherently interpretable model architectures and their diagnostics.
Locating failure regions instead of averaging over them.
Knowledge-graph embeddings as geometric operators; the mathematics behind GMS.
Deterministic policy enforcement around probabilistic reasoning.
Design of experiments, graph-derived ground truth, judge-reliability measurement.
Book-length treatments and hands-on workshop material for model risk and engineering teams.
Benchmark results are published with datasets, baselines and reproduction code as they complete our claims-evidence process.
