NereidAI

Navigate the sea of data — from devices, to models, to edge intelligence.

An end-to-end, air-gapped, multi-silicon AIoT platform. IoT connectivity and AI intelligence in one closed loop — private by design.

Explore the suite Architecture

Open-source status — now open, in active development

Every NereidAI repository is now open source on GitHub under Apache-2.0, having passed compliance, open-source standard and completeness review. The code currently ships on its development (dev) channel — there is no tagged release yet. Commercial terms, support SLAs and enterprise deliverables will be documented here when the first formal release is published. Browse the repositories at github.com/NereidAI and track progress on the roadmap.

The closed loop

NereidAI ships as a suite of focused products; each can be deployed standalone, and together they form one unified platform. A thin suite layer ties them together without absorbing their domain logic.

flowchart LR
    subgraph FIELD["Site / field"]
        DEV["cameras · sensors · robots"] --> EDGE["NE · Edge\ncollect & bridge data"]
        DEV --> NIN["NIN · Inference\nrun model inference on the gateway"]
    end
    FORGE["NF · Forge — the model factory\nannotate → train/fine-tune → quantize →\nmulti-silicon export → evaluate → register"]
    HORIZON["NH · Horizon\nmodel & app rollout (OTA)"]
    LINK["NL · Link\ningest · rules · alerts · dashboards"]

    EDGE -->|data / feedback| FORGE
    LINK -->|datasets & telemetry| FORGE
    FORGE -->|versioned models| HORIZON
    HORIZON -->|OTA models / apps| EDGE
    HORIZON -->|model packages| NIN
    NIN -->|inference events| LINK
    EDGE -->|telemetry| LINK

The loop has two halves that meet in Forge: the IoT half brings field data in (Edge, Link) and acts on results; the AI half turns that data into models (Forge) and ships them back to run at the edge (Horizon, Edge, Inference). Models improve from real telemetry, on the next iteration — with no mandatory cloud.

• Private by design

  ---

  Fully offline / air-gapped delivery with no mandatory call-home. Data, training and inference all stay inside the customer perimeter.

• From CNNs to large models

  ---

  One factory for classic vision models (detection / classification / segmentation) and large models — VLM-assisted labeling plus LLM / VLM fine-tuning and serving.

• Multi-silicon → edge inference

  ---

  Export one trained model across NVIDIA, AMD, Intel, ARM, Ascend, SOPHGO, MetaX, Enflame and Rockchip, then run low-latency CV or VLM inference on the gateway.

• Massive device access, in real time

  ---

  Multi-protocol ingress (MQTT / HTTP / CoAP), a rule engine, alerting and dashboards — every change pushed over WebSocket, no polling.

• Edge autonomy on flaky links

  ---

  Field gateways keep working while the WAN is down: local autonomy plus store-and-forward buffering that safely replays telemetry on reconnect.

• Offline distribution at scale

  ---

  NereidHorizon rolls out Edge itself, apps and models to thousands of remote nodes — policy-driven, bandwidth-aware OTA with offline reconciliation.

• One AIoT closed loop

  ---

  Field data → models → OTA rollout → edge inference → telemetry feedback: IoT connectivity and AI intelligence in a single loop, no mandatory cloud.

Apache-2.0 community edition · enterprise LTS, SLAs, hardening and turnkey private delivery — see Community vs Enterprise.

From data to edge intelligence

The AI journey is a first-class path, not a side feature:

#	Stage	What happens	Product
1	Collect	Cameras, sensors and robots feed images, video and time-series data from the field.	Edge, Link
2	Annotate	Datasets are versioned; vision-language models (Florence-2, Qwen-VL) pre-label for human review.	Forge
3	Train / fine-tune	Kubeflow pipelines train classic vision models or fine-tune LLMs / VLMs (LLaMA-Factory) on the customer's own data.	Forge
4	Optimize & export	Quantize / compile (NNCF, OpenVINO, TensorRT, vLLM) and export for the target silicon.	Forge
5	Evaluate & register	Quality and benchmark metrics — including LLM-as-a-judge — gate each version; approved models enter the registry.	Forge
6	Run at the edge	Models roll out over the air and run as low-latency CV or VLM inference on gateways and devices.	Horizon, Edge
7	Monitor & feed back	Inference events and telemetry drive rules, alerts and the next training round.	Link, Forge

The same loop powers the commercial solutions — smart-manufacturing inspection, video/image analytics (NereidSentinel) and medical insights (NereidMedical).

Product suite

Two thin suite services tie the family together; five domain products do the work. Every domain product is a complete system that can run on its own. Short names are used in prose and abbreviations in diagrams — see Naming & abbreviations.

Product	Abbr.	What it does
NereidSphere	NS	Thin suite layer — single portal, unified API gateway, and cross-product event glue.
NereidIdentity	NI	Suite SSO and identity center — OIDC / LDAP / SAML federation, RBAC, token issuance.
NereidLink	NL	Privatized IoT data pipeline — device access, rule engine, alerting, dashboards.
NereidForge	NF	Model factory — datasets, training, optimization, multi-silicon export, registry.
NereidHorizon	NH	Edge distribution — policy-driven rollout, bandwidth-aware delivery, OTA.
NereidEdge	NE	Southbound edge integration on field gateways — zero-touch, mTLS, autonomy.
NereidInference	NIN	Edge inference runtime — camera frames to model servers over KServe/OpenAI, structured events to the EdgeX bus.

Where to start

• Getting started

  ---

  Install the community edition, run your first loop, and connect a device.

  Development

• Deployment

  ---

  Air-gapped installation, Kubernetes & Helm, and edge delivery.

  Deployment

• Security

  ---

  SSO / OIDC, RBAC, mTLS, and compliance posture.

  Security

• Community vs Enterprise

  ---

  What is open source, and what is offered under commercial terms.

  Compare