KnowlytiX builds the governed harness that makes agentic AI trustworthy enough for production in regulated industries. The team's premise is simple: AI compliance should be mathematical, not probabilistic. Retrieval-augmented generation solved the problem of finding the right document, and guardrails solved the problem of filtering unsafe content, but neither verifies that a model's answer actually complies with domain-specific rules. Knowlytix builds that missing layer — extracting claims from an agent's output, verifying each one against structured domain rules, and producing auditable decision trails that hold up under regulatory scrutiny.

At the center of that layer is the Geometric Memory System (GMS), a deterministic substrate that replaces the guesswork in a typical LLM pipeline with components that produce evidence. Where a prompt-and-pray agent relies on instructions a model can ignore, a flat memory store that cannot tell a fact from a fabrication, and a second model grading the first, GMS substitutes a calibrated admissibility gate, a structured store that admits a fact only after it passes the same check an action would, and an independent, non-LLM verifier whose verdicts are byte-exact and reproducible. GMS reached a stable 1.0 release on PyPI in July 2026, and core components — including the agentlab library and the Knowledge Adapter Layer (KAL) — are open-sourced under Apache 2.0.

Knowlytix publishes its research as it goes, on the view that a claim to trustworthy AI is only as credible as the evidence behind it. That work includes FinStructBench, a benchmark for structured information retrieval from financial documents; a study measuring where LLM-as-judge evaluation disagrees with graph-verified ground truth; a GPU-accelerated design-of-experiments method for systematically evaluating language models on regulatory documents; and KAL, a specification for connecting knowledge graphs to geometric memory systems. This book series grew directly out of that research and out of the engineering work of building GMS itself — it is the practical, code-first counterpart to the papers, written for engineers who need to build agentic and retrieval systems that a compliance team, an auditor or a regulator can actually examine.