Skip to content

NereidForge

Forge is a privatized model factory: an on-premises platform that turns business data — datasets, images and video — into models that run on field devices, covering data and annotation management, training and fine-tuning, optimization and quantization, hardware-targeted export, evaluation and a model registry.

What it solves

  • Data cannot leave the organisation. Industrial, government and security data is constrained by regulation; the whole factory runs inside the customer's own environment and works fully offline, with base models shipped rather than downloaded.
  • AI projects are hand-crafted. Data preparation, annotation, training, evaluation and deployment are stitched together by a few experts across many separate tools; Forge turns that into one repeatable production line.
  • Capability lives in individuals, not the organisation. Data, models and evaluation reports are registered and versioned, so knowledge outlives staff turnover.
  • Lab models do not reach the field. Field devices have limited compute and many silicon variants; one training run produces models targeted at multiple hardware routes.
  • Vendor and cloud lock-in. The platform is privatized and open in its data and format choices, so today's investment does not become tomorrow's island.

Core capabilities

Data and annotation

  • Dataset management with versions and traceable provenance, including draft-to-final workflows.
  • Multi-format import — object detection, classification, segmentation and instruction formats — with a canonical, split-preserving layout.
  • Annotation through integrated Label Studio and CVAT studios, covering images and video.
  • Machine pre-annotation: a model proposes boxes so annotators review and correct rather than draw from scratch; video annotators label key frames and the platform tracks the frames between.
  • Per-project assignees and collaborative review.

Training and fine-tuning

  • A standardized pipeline an ordinary engineer can operate, from dataset to trained model.
  • Support for computer-vision models and for language / vision-language models, with fine-tuning on organisation data.
  • Per-vendor training images keep vendor-specific acceleration and toolchains isolated; there is never a single catch-all image.
  • A small Trainer SDK (Python) defines one environment-in/artifacts-out contract for custom trainer images, so partner or in-house frameworks run in their own image and registry; the built-in trainers use the same SDK.

Optimization, export and evaluation

  • Optimization and quantization plugins (compile / quantize) applied before a model is released.
  • Multi-framework export to a portable runtime format, with automatic detection of the source framework family.
  • Hardware-targeted export, including NVIDIA, AMD, Intel and domestic silicon routes, so hardware procurement keeps its options.
  • Automated evaluation before release, with the report kept alongside the model version.

Registry, serving and verification

  • A model registry that answers, for any version, what data trained it, how it scored and where it is deployed.
  • Model serving with on-demand loading of uploaded weights.
  • "Try with image" verification of any model artifact without deploying first, with a visualization of what the model attends to.

Delivery

  • Fully privatized delivery: images come from an internal registry and the stack runs in an offline machine room.
  • Declarative deployment and upgrade on Kubernetes.

How it fits in the suite

Forge is the model-factory stage of the closed loop. It consumes data collected at the edge — the telemetry Link lands — and produces the models and applications that Horizon distributes to edge nodes. Per the suite's design, vendor-specific acceleration, images and toolchains are isolated inside Forge pipelines; every other product consumes versioned artifacts, never vendor tooling.

flowchart LR
    LINK["NL"] -->|telemetry data| FORGE["NF"]
    FORGE -->|models / apps| HORIZON["NH"]
    HORIZON -->|OTA| EDGE["NE"]
    EDGE -->|data| LINK

Developer documentation

The authoritative, detailed developer documentation for Forge lives in its own repository: https://github.com/NereidAI/NereidForge.

Key facts

Aspect Detail
Role Privatized model factory — datasets, training, optimization, export, evaluation, registry
Value Turns business data into models that run on field devices; assets stay with the customer
Annotation Integrated Label Studio and CVAT studios with machine pre-annotation
Model types Computer-vision models and language / vision-language models
Export Multi-framework export to a portable runtime format, targeting several hardware routes
Serving Model registry plus model serving with try-with-image verification
Private by design Runs fully offline in the customer's own environment; base models shipped with the system
Delivery form Declarative Kubernetes deployment from an internal image registry
Language / stack Python backend and React frontend; orchestrated pipelines and model serving on Kubernetes
License Apache-2.0

Resources