01 · Speed
Sub-10ms Inference
Run vision, anomaly, classification and movement models locally with sub-10ms latency — fast enough for safety-critical decisions.
A lightweight execution layer that runs directly on commodity x86 and ARM gateways. INNFINI processes sensor data, applies rules, runs local inference and triggers actions at the site — even before signals reach the central platform.
Six stages. INNFINI connects locally, filters the noise, runs inference on the gateway, applies your rules, buffers through outages, and syncs clean events back when the network's there.
RFID readers, cameras, BLE gateways, sensors, PLCs, BMS controllers — all wired directly to the edge node, no cloud detour.
Dedupe reads, cleanse signals, aggregate streams. 12,847/s raw becomes a few dozen meaningful events.
Vision, anomaly, classification, movement direction — fast local inference on x86 or ARM, no GPU required.
Apply thresholds, business rules and decision logic at the gateway — alarms, escalations and actions fire instantly.
Store events locally during network drops. Continue local alerts. Preserve timestamps. Resume the moment uplink returns.
Clean, structured, time-stamped events stream securely back to INNFINI — feeding the graph, agents and dashboards.
Fast where speed matters. Resilient where uptime matters. Cheap where the math matters. No proprietary boxes, no GPU mandates, no vendor cages.
Run vision, anomaly, classification and movement models locally with sub-10ms latency — fast enough for safety-critical decisions.
Deploy on commodity gateways, industrial PCs, mini PCs, ruggedized field devices and edge servers. No proprietary boxes required.
Keep capturing, processing and storing events through network drops. Resume sync transparently the moment uplink returns.
Apply thresholds, validations, filters and business logic at the edge — the same rules language that runs in the cloud.
Filter, aggregate and reduce traffic upstream. Send only the events that matter — 98% bandwidth saved on noisy sites.
mTLS-encrypted, signed and audited handoff back to INNFINI cloud. Every event has provenance from the gateway forward.
The runtime is a single signed binary with no kernel modules. Sizing follows the workload, not a vendor SKU — these are the tiers we deploy against, and the workload each one carries in production.
| Tier | Typical hardware | Footprint | Carries | Power |
|---|---|---|---|---|
| Micro | ARM64 gateway · 4 core · 4 GB | 210 MB RAM · 1.2 GB disk | Rules + filtering, up to 40 sensor streams | < 10 W |
| Standard | x86 industrial PC · 8 core · 16 GB | 1.4 GB RAM · 40 GB disk | 4 vision models, 8 camera streams, 500 sensors | 35–65 W |
| Site | Edge server · 16 core · 64 GB · T4 or A2 GPU | 6 GB RAM · 500 GB disk | 24 camera streams, full local graph, 30-day buffer | 180–300 W |
| Rugged / mobile | Fanless IP65 · ARM64 or x86 · DIN or vehicle mount | 380 MB RAM · 8 GB disk | Vehicle telematics, mobile command, field kits | 12–24 V DC |
Sub-10 ms is a budget, not a slogan. Measured on the Standard tier for a 1080p vision detection, frame arrival to action emitted — the same path a cloud round trip covers in 380–600 ms.
| Step | p50 | p99 | Notes |
|---|---|---|---|
| Frame capture and decode | 1.8 ms | 3.1 ms | Hardware decode where available |
| Pre-processing and tensor prep | 0.9 ms | 1.6 ms | Zero-copy into the accelerator |
| Model inference | 3.4 ms | 6.2 ms | INT8 quantised, batch of one |
| Rule evaluation | 0.6 ms | 1.1 ms | Same rule language as cloud |
| Action emit and local persist | 1.1 ms | 2.4 ms | Written before acknowledgement |
| End to end | 7.8 ms | 14.4 ms | Cloud round trip: 380–600 ms |
Every edge platform claims offline support. The honest question is which capabilities degrade, and in what order — so here is the full answer, including what does not survive.
| Capability | Disconnected | Why |
|---|---|---|
| Sensor and video ingest | Full | Nothing in the capture path depends on the uplink. |
| Local inference | Full | Models are resident on the node, not called remotely. |
| Rules and local automation | Full | Cached rule set continues; changes queue for the reconnect. |
| Local alerting and signage | Full | On-site outputs are driven from the node itself. |
| Cross-site correlation | Degraded | The node can only correlate what it can see. Multi-site patterns resume on reconnect. |
| Graph enrichment | Degraded | Falls back to the cached subgraph for local assets; global lookups defer. |
| Consequential automation | Held | Deliberate: actions needing central authority are queued for approval, never taken unilaterally. |
| Model retraining and promotion | Suspended | Training is a central function. The node runs its current model unchanged. |
The difficulty of edge is not the first node — it is updating the four hundredth without a truck roll, and knowing which eleven are unhealthy right now.
Updates roll canary → ring → fleet with health gates between stages. A failed health check halts the wave automatically rather than continuing into the estate.
The node keeps the previous image on a second partition. A failed boot or failed health check reverts locally without anyone visiting the site.
Every node's state is declarative and versioned. A replacement box is provisioned by identity and converges to the correct configuration on first boot.
Temperature, disk wear, model confidence drift, buffer depth and uplink quality stream back as first-class telemetry — a degrading node raises an incident before it fails.
Edge hardware sits in corridors, cabinets and vehicles. The threat model assumes someone can touch it.
| Control | Implementation |
|---|---|
| Measured boot | TPM 2.0 or ARM TrustZone verifies the boot chain. A node that fails attestation is refused enrolment and cannot rejoin the fleet. |
| Encryption at rest | Full-disk encryption with the key sealed to the TPM, so a removed drive is unreadable off the node. |
| Mutual TLS identity | Each node holds a unique hardware-bound certificate. Certificates rotate automatically and can be revoked centrally within seconds. |
| Signed artefacts | Runtime images, models and rule sets are signature-verified before load. An unsigned or altered artefact will not execute. |
| No inbound ports | The node dials out. There is no listening management port to scan, and no inbound path from the network into the runtime. |
| Tamper response | Chassis intrusion and unexpected reboots raise a security incident, and the node can be configured to seal its keys until re-attested. |
| Data minimisation | Video can be processed to events on-node and never transmitted, so imagery of people need not leave the site at all. |
Actions fire locally instead of round-tripping through the cloud — critical for safety, security and high-throughput operations.
Sites keep scanning, dispatching, alerting and recording through network drops, brownouts and ISP outages.
Send clean operational events, not continuous raw streams — cuts upstream bandwidth on noisy RFID and vision sites.
x86 IPC, ARM gateway, rugged field unit, edge server, mini PC — pick the form factor that fits each site.
The Edge Runtime extends platform intelligence to the physical site — enabling faster decisions, resilient operations and real-time automation even when connectivity is limited.
A 60-minute architecture review with our solutions team. We map your sites, your hardware, and the workflows where edge processing changes the math.