ModelForge

Deploy any model. Operate it safely.
A model registry, deployment control plane, and inference platform built for
versioned artifacts, progressive releases, and framework-independent execution.

What it does · Product preview · Quick start · Architecture · Documentation

---

What it does

ModelForge gives teams one place to register models, publish immutable versions, route inference traffic, and manage deployments. The same control plane serves browser, CLI, and API workflows.

Capability	In practice
Version with confidence	Publish immutable artifacts and verify their SHA-256 checksums before execution.
Release progressively	Send a weighted share of requests to a canary, then promote, abort, or roll back.
Run across frameworks	Use PyTorch, ONNX, the Go runtime, or an independently implemented runtime plugin.
Contain failures	Inspect health and metrics; use bounded retries, circuit breakers, and stable fallback.
Separate workspaces	Scope models, deployments, artifacts, and API keys to the authenticated workspace.

Product preview

The browser console surfaces model versions, serving targets, runtime health, and release actions from the same API used by the CLI.

_{Illustrative console view based on the current interface, using example data. ModelForge is not publicly hosted yet.}

A release in one view

The stable model continues serving while a candidate receives a controlled share of traffic. The scenario can promote the candidate or restore the stable target after a rollback or failed canary.

The deployment validation scenario exercises this flow through real API requests and records the results as JSON and Markdown evidence.

Quick start

Want to look around first? View the illustrated product preview. The application is not publicly hosted yet.

To run the local development stack, install Docker with Compose, then run:

git clone https://github.com/michaelbawuah/ModelForge.git
cd ModelForge
docker compose up --build --wait --detach

When the containers are healthy, open http://localhost:8000/ (landing page) or http://localhost:8000/dashboard (console) on the same computer. These addresses only work while your local stack is running; they are not links to an online instance.

To run the optional deployment scenario, install Python 3.11+ and the project dependencies in a local environment, then run:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
make proof

Inspect demo-proof-evidence/proof-report.md and stop the stack when finished:

docker compose down

The Compose configuration uses development defaults. Follow the cloud deployment contract for a hosted installation; the AWS reference stack has not been launched for public use.

Architecture

flowchart TD
    U["Console · CLI · API"] --> C["FastAPI control plane"]
    C --> D["Model registry · deployment state"]
    C --> A["Verified artifact storage"]
    C --> R["Runtime resolver"]
    R --> P["PyTorch · ONNX · plugins"]
    R --> G["External Go runtime"]
    P --> O["Prediction · metrics"]
    G --> O

Loading

The control plane owns model identity, workspace boundaries, release state, and traffic decisions. Runtimes own framework-specific model execution. New frameworks can implement the external runtime contract without adding framework-specific branches to the serving path.

Built with

Control plane	Execution	Data and operations
Python · FastAPI · SQLAlchemy	PyTorch · ONNX Runtime · Go	MySQL · Alembic · Prometheus
OIDC browser sessions · workspace API keys	Runtime plugin protocol · HTTP	Local/S3-compatible artifacts · Docker · Terraform

CI runs Python and Go validation, Terraform checks, container builds, and a Compose smoke test. Benchmark figures are documented in the engineering guide with their CI hardware context; they are not production capacity claims.

Documentation

• Platform overview — lifecycle, runtimes, and access model.
• Engineering guide — implementation details, commands, tests, and measurements.
• Architecture and invariants — serving and failure boundaries.
• Deployment validation — promotion, rollback, and regression checks.
• Runtime protocol — add an external framework runtime.
• AWS reference deployment — infrastructure design and release workflow.

Status

ModelForge runs locally and has a tested AWS deployment reference. Public hosting, production identity configuration, and live AWS verification remain future work. The repository's CI and demo evidence describe tested behavior; they do not imply a running public service.