frontedge

The frontend layer for ML engineering: a ready-made design system and screens that mount into any FastAPI application.

pip install frontedge

The interface is inside the wheel; no Node build is needed. The design system for your own screens is @frontedge-ml/ui on npm.

from fastapi import FastAPI
import frontedge

app = FastAPI()
app.include_router(frontedge.runs(runs_provider=my_provider))

The provider returns frontedge.Run objects; the library knows nothing about authentication and hands the request to the provider, so access control stays on the application side.

What's inside

• frontedge.mount(runs_provider, evals=..., datasets=..., actions=..., config=..., path="/frontedge") — the router: the list of runs, the run details, the log stream (SSE), the metadata and the built interface.
• frontedge.runs(runs_provider, path="/frontedge") — the same router without the eval and dataset screens.
• frontedge.models — the domain model: Run, Stage, MetricPoint, Artifact, Example, LogLine, EvalSummary, EvalComparison, DatasetInfo, DatasetDetail, the statuses and the freshness.
• frontedge gen --out src/generated — the generator of TypeScript types and a typed client from the pydantic models. Pure Python, no Node needed.

One process, one port: the interface and the API are served by the same app.

Configuration

The configuration is data, not code — a TOML or JSON file that the same interface can edit by hand, from code and later with a visual editor:

# frontedge.toml
density = "default"          # compact | default | comfortable
theme = "light"              # light | dark
run_columns = ["name", "status", "ribbon", "dataset", "startedAt", "durationMs", "accuracy", "loss"]

app.include_router(frontedge.mount(my_runs_provider, config="frontedge.toml"))

UiConfig also knows eval_tabs, actions, live_refresh_seconds, baseline_run and page_size. The configuration is served to the browser as GET /frontedge/api/config.

Actions

The interface never invents what a button does: a project declares the actions it supports, and the library calls them.

[[actions]]
id = "stop"
label = "Stop"
placement = "run"        # run | runs | evals | datasets
danger = true

from frontedge import ActionResult

async def stop_run(request, payload) -> ActionResult:
    run_id = payload["runId"]
    await my_repository.stop(run_id)
    return ActionResult(ok=True, message=f"Run {run_id} stopped")

app.include_router(
    frontedge.mount(my_runs_provider, actions={"stop": stop_run}, config="frontedge.toml")
)

The handler receives the request and the payload of the interface and returns an ActionResult; its message is shown to the user.

The full description of the product and of the visual language: design document. The full contract: docs/integration.md.

MIT.