Nabu puts verifiable AI behind every investigation: every claim cited, every step recorded, from first byte to final claim. It reads, organises, and cross-references evidence at scale — documents, records, and data — and returns findings a professional can verify, with the source behind every statement and a tamper-evident record of every step in between.
AI does the reading; people make the findings. Nabu drafts, retrieves, and cites; a qualified human reviews and commits every conclusion that matters. That line never moves.
This line holds in every Enkidu sector — see Trust.
Nabu serves the legal, law-enforcement, and intelligence sectors through implementations tuned to each domain — announced as they enter service.
One Nabu, bounded by zone
Nabu deploys as multiple cooperating instances, each placed inside a different security zone. Every instance is wrapped in its own Anshar zone — the configured horizon that determines what that instance can reach. The central instance holds the case data and never reaches outside; a separate instance in the DMZ brokers external services and never touches the case data.
Built on Ekur
Nabu runs on Ekur, the shared technology stack beneath everything we build. Its capabilities are not products — they are the common foundation that makes every Enkidu platform secure, auditable, and sovereign by construction.
Sectors
Nabu’s reference deployment runs a reasoning model, retrieval and embedding models, a speech model and a safety-screening model as co-resident services — each an open-weight base adapted for the domain, each a separate custody item that can be adapted, replaced or rolled back on its own. Enkidu does not pre-train foundation models; it makes strong open bases excellent at your task, and commits to keeping fielded models current under the zero-regression rule, redistributed under Nusku. The operator declares which adaptation each continuum uses.
A Nabu implementation is customised for a client by selection, not by rewrite. Ummânu — the Enterprise AI Workforce Instrument — supplies the roles: 184 specified personas with sources, gates and a refusal rule, staffing the implementation’s agents. Kulla — the Governed Skill Production Instrument — supplies the checks: promoted skills that execute as deterministic gates, with their exit codes recorded in custody. Both are stamped into the custody record of everything they touch; neither self-instantiates nor self-promotes.
Enkidu is evaluating NVIDIA ModelExpress for scaled implementations of Nabu — infrastructure that moves model weights peer-to-peer across a fleet, wrapped in Enkidu integration patterns that keep every transfer inside governance: one governed entry point, a signed-lineage install gate, and zone-scoped distribution.
Evaluation and prototyping activity; no availability or performance claims.