Home/Sectors/Urban
Sector · Urban

Lamassu

Lamassu — the Urban and Public Safety implementation of Baru, the trusted intelligence-fusion platform. The guardian at the gate — ambient sensing for the city, with evidence behind every alert and a human behind every response.

It raises the alert. It never takes the action.

Urban
2 tiers
Edge nodes · one central node
5 modalities
Vision · acoustic · RF · spatial · environmental
1–8B
Sensor-tuned micro-models at the edge
Minutes
Node deployment to operational
The implementation

Determinism where it matters, reasoning where it helps

Lamassu drops self-contained edge nodes into a neighbourhood, a campus, or a whole city. Each node connects to the sensor feeds it finds locally — cameras, acoustic arrays, air-quality networks, building systems — runs sensor-specialist micro-models on what it sees, and shares observations with its neighbours over a self-healing mesh. One central node correlates the whole fabric and commits the picture an operator acts on.

A node is productive within minutes of mounting: it self-provisions with a cryptographic identity, discovers local feeds over the standards they already speak, and keeps sensing on its own when the network fails — reconciling when the link returns.

    What it delivers
  • Deployable nodes, operational in minutes
  • Ingestion over the standards the city already speaks
  • A picture that survives the loss of any node or link
  • Confidence computed from evidence, not asserted
  • Privacy-first: Personal Data redacted, retention-limited
The fabric

A mesh of nodes, one centre

Deliberately two tiers — an urban deployment doesn't need a forward-fusion hierarchy. Edge nodes are mesh peers to their neighbours and spokes to a single central node sized for the city.

7–25 W · 20–67 TOPS

Compact Node

A pole- or wall-mounted single-modality pod — one or two streams, minimal local inference, a full mesh peer. Runs on mains, PoE, or battery with solar.

10–40 W · 70–157 TOPS

Standard Node

A multi-sensor pod or cabinet running concurrent perception and small on-board models — the workhorse of the fleet, mesh peer and spoke at once.

15–60 W · 200–275 TOPS

Heavy Node

Multi-camera, multi-modal fusion with larger local models — and a local sub-aggregator for a cluster of lighter nodes where bandwidth is tight.

Server-class · one per city

Central Node

Cross-node correlation, the coordinator model, the adjudication engine, long-term stores, and the operator console — the only tier that commits the picture.

Speaks the city's standards ONVIF / RTSP MQTT OGC SensorThings NGSI-LD OPC-UA / Modbus DDS CAP alerts
The intelligence model

Propose. Compose. Commit.

Many small experts instead of one large model — a bet that specialisation plus composition beats generalisation for urban sensing. And a three-step transition no model may short-circuit.

01 · At the edge

Specialists propose

One micro-model per modality — vision, acoustic, RF, spatial, environmental — each reading only its own telemetry and emitting cited, typed findings. On out-of-distribution input it abstains rather than guessing.

02 · At the centre

The coordinator composes

A single larger reasoning model plans the run, dispatches work to specialists across nodes, and composes their findings across modalities into a candidate event with its provenance. It never assigns the verdict.

03 · Deterministic

The engine commits

Deterministic, code-enforced logic — not any model — converts proposals into the committed, operator-facing indicator under bound policy. Reproducible, auditable, and resolvable to signed provenance through a full trace tree.

Many missions

A mission is defined, not built

The same fabric hosts many missions at once, each an isolated Anshar execution zone under its own policies and sensitivity ceiling. Adding one is a configuration act under governance — not a re-architecture.

Critical infrastructure

A standing watch around a utility, substation, water plant, or port — approaches classified and raised before they reach the boundary.

Environmental early warning

Air-quality, gas, flood, and wildfire-smoke sensing fused into early warnings — confidence rising as independent sensors corroborate.

Public-safety awareness

Crowd density, flow, and anomaly sensing at events and transport hubs — with redaction and retention limits applied to personal data by default.

Airspace safety

Unauthorised drones over airports and stadiums, detected by fusing RF, acoustic, and visual modalities — response owned by the responsible authority.

The line that never moves

It raises the alert.
It never takes the action.

Lamassu is civilian, without exception. The most consequential thing an agent may do unaided is raise an alert; every effect that reaches the outside world — even steering a camera — is gated behind a deny-by-default policy engine, an earned trust zone, and an accountable human decision. The policy layer of every zone is human: systems do not make decisions for action, because they cannot be responsible for them.

This line holds in every Enkidu sector — see Trust.

Governed for civil deployment
Data sensitivityPublic · Internal · Sensitive · Personal
Privacy & securityGDPR · ISO 27001 / 27701
AI managementISO 42001 · NIST AI RMF
Agent authorityTrust zones Z0–Z3 · deny-by-default
External alertsOASIS CAP · operator-owned
Lineage

Part of Baru

Lamassu is Baru, tuned for the city: the same platform, carrying urban-sensing models, municipal policy packs, and public-accountability posture. What Baru guarantees everywhere — evidence behind every element, a human behind every decision — Lamassu guarantees on the street.

The foundation

Built on Ekur

Every node in the mesh runs the same shared stack — one layer for where things may happen, one for whether they may. Anshar bounds each operation to a zone with explicit entry and exit conditions; Kittu assigns trust zones to agents, gates every sensor and actuator, and captures the chain of custody behind every indicator. The full detail lives on the platform page.

Enkidu

Bring Lamassu to your city

Book a walkthrough and see a node go from the crate to an operational mesh peer in minutes — on your own feeds.