Enlil — the Defence implementation of Baru, the trusted intelligence-fusion platform. Fusion for the tactical edge — sovereign, disconnected-capable, and always under human command.
Enlil gives a forward unit the fused picture and decision support that today only a rear Joint Intelligence Centre can provide — on hardware a platoon can carry. It is an intelligence-fusion platform, not a combat platform: each node infers on what it can see, shares that view with its neighbours and hub, and reconciles upward through a forward fusion centre to the rear — one continuous fabric from the smallest node to the data centre.
The inference that decides the next minute runs on the node, not in the rear. Reachback enriches; it is never a dependency. Any node, hub, or link may be lost and the surviving fabric continues to observe, infer, and inform.
A mesh gives peer-to-peer resilience; a distributed hub-and-spoke overlay gives the inference hierarchy. Every tier maps to NVIDIA silicon sized to its role — so weight, power, and inference capacity match the job.
Fuses forward. Reconciles upward.
Higher-confidence inference over a cluster.
Dismounted operators, vehicles, autonomous platforms, emplaced sensors.
If the hierarchy is cut, the mesh keeps the picture alive — nodes operate autonomously and reconcile when the link returns.
Four tiers. At tier one, the edge, nine fusion nodes — dismounted operators, vehicles, autonomous platforms and emplaced sensors — are joined to their neighbours by a peer-to-peer mesh, so no single node is a point of failure. Each node also connects upward by a heavier spoke to one of three tier-two cluster hubs, which perform local fusion: higher-confidence inference over a cluster. The three hubs report to a tier-three forward fusion centre, the FOB-IFC, which correlates across hubs. That reports to the tier-four Central IFC, responsible for national all-source fusion and direction. Observations and assessments flow upward; direction flows back down. If the hierarchy is cut, the mesh keeps the picture alive — nodes operate autonomously and reconcile when the link returns.
Carried by a dismounted team, a crewed vehicle, or an autonomous platform (UAS, UGV, USV). Runs limited on-board inference and exchanges observations peer-to-peer across the mesh.
A higher-capacity unit acting as a fusion centre for a cluster of local nodes — running the harder inference a node abstains to, and hosting the seam where local fusion meets the unit's C2.
A forward fusion centre at the Forward Operating Base. Connects dozens of hubs and runs models up to ~80B parameters for forward all-source correlation, then reaches back to the rear.
The rear, data-centre-class tier. Supplies the deepest all-source reasoning with Large and Frontier models, and returns national direction and feeds down the same path.
The silicon rationale for every tier — read the engineering note →
Four stages left to right, all recorded on one continuous chain of custody that is signed, hash-linked and tamper-evident. Ingestion is a four-step pipeline — authenticate, tag and classify, correlate, record custody — and nothing moves untagged, unauthenticated or unrecorded, under any load. Observation is the common tagged object carrying identity, origin, classification and custody. Track is correlated observations of one entity. Assessment is inference over tracks with lineage pointers back to every parent. Each stage drops a tick onto the custody band. The ladder ends at a separate block, set apart by a gap and its own border: a human decision, outside the automated flow.
Some are sensed forward by the nodes and hubs themselves; others arrive from the Central fusion centre over reachback. All enter through one common ingestion pipeline that tags classification at acquisition and checks it at every share.
COMINT, ELINT, direction-finding and electronic order of battle.
EO/IR imagery from organic RPAS forward; satellite imagery from the centre.
Acoustic, seismic, and RF measurement signatures from emplaced sensors.
SALUTE and spot reports, source reliability, document exploitation.
Foundation maps, terrain, line-of-sight and mobility analysis — field-updated.
Open-source reporting and context, brokered from the enterprise.
Staff product and direction from CJOC J-2, pushed down to the edge.
Five Eyes and coalition feeds, brokered centrally for releasability.
Enlil's actuator authority reaches only to intelligence collection — tasking, repositioning, or steering a sensor to gather data — which an agent may issue autonomously once it has earned a high trust zone. It never authorises a kinetic effect. An engagement is always cued to, and decided by, a cleared human operator or the combat platform's own function.
This line holds in every Enkidu sector — see Trust.
Enlil is Baru, tuned for Defence: the same platform, carrying defence-grade models, policy packs, and compliance posture. What Baru guarantees everywhere — evidence behind every element, a human behind every decision — Enlil guarantees at the tactical edge.
Every tier of the fabric runs the same shared stack — one layer for where things may happen, one for whether they may. Anshar bounds each operation to a physical zone — a sector of terrain — and a logical continuum with explicit entry and exit conditions, then tracks the movement of threats and notifies exactly the nodes, hubs, and fusion centres that need to know. Kittu, a trust and security layer over NVIDIA NeMo Guardrails, assigns trust zones to agents, controls sensor and actuator use, and captures the end-to-end chain of custody behind every inference — the evidence that becomes confidence in a threat. The full detail lives on the platform page.
Book a technical briefing and see fusion, zoning, and the trust harness running on ruggedised edge hardware.
Enkidu is evaluating a connected set of technologies within Enlil’s fabric: NVIDIA Halos — the Holoscan Sensor Bridge as a trusted sensor ingestion path, NVIDIA IGX Thor at the hub tier as a deterministic safety authority, and the Halos Safety Core pattern as a Kittu runtime pattern; and, for efficient reasoning at the edge, quantization-aware distillation paired with speculative decoding — compressed reasoning models whose fidelity to their full-precision parents is measured, recorded, and treated as assurance evidence.
Evaluation and prototyping activity; no availability or performance claims.