PLATFORM

Three layers, each able to run without the one above it

A wristband keeps evaluating rules when the gateway is unreachable. The gateway keeps holding approvals when the orchestrator is unreachable. Nothing about the design assumes the network is there.

THE STACK

Every layer, end to end

Firmware

MICROCONTROLLER CLASS

Real-time scheduling with a verified boot chain, so a device that cannot prove what it is running cannot influence a rule.

FreeRTOSTensorFlow Lite MicroCMSIS-DSPMCUbootSecure element

Edge

GATEWAY

Streaming ingestion, deterministic rule evaluation, and an approval queue that survives a power cut.

Rust servicesKafkaLocal DAG engineSQLite WAL queuemTLS

Orchestration

SERVER

Durable workflows survive restarts and partial failures. A run resumes rather than replays.

TemporalTimescaleDBPostgreSQLObject storePolicy engine

Explanation

MODEL LAYER

Turns scores and signal features into the sentence at the top of an approval card, constrained to evidence present in the proposal.

Rule provenance retrievalTemplated groundingNo free-form claims

Applications

WEB AND MOBILE

Operations console, rule builder, fleet view and ledger. The approval surface is designed for a phone at arm's length, one-handed.

Next.jsReactWebSocket streamsFIDO2 passkeys

Identity

CROSS-CUTTING

Workload identity everywhere, human identity at the gate, and a migration path for long-lived signatures.

SPIFFE/SPIREFIDO2Hash-chained ledgerPost-quantum roadmap

Bounded evaluation

Rules compile to a directed acyclic graph. Evaluation is side-effect free, time-bounded and replayable against recorded signal.

Offline approvals count

Answers given during an outage are signed locally and reconciled when the link returns, never dropped, never replayed twice.

Time you can defend

Devices are time-disciplined against the gateway, so a fused event carries an ordering that stands up in an incident review.

Designed for DDIL conditions: degraded, disconnected, intermittent and limited bandwidth.

Deployment models

Start with one ward, one shift, or one squad

A pilot runs four to sixteen weeks depending on deployment model. You bring the population and the responders. We bring devices, gateways and the first set of rules, and we measure override rate and time-to-answer from day one.

Request a pilotSee how it works