Talk to sales
Workbook

Machine Learning Engineer

You own: model training, evaluation, registry hygiene, and shipping models to production with GeneFlow.

What's new for you

WasNow in GeneFlow
mlflow.start_run() with no $ trackinggeneflow.start_run() — same shape, plus cost_usd per run
Hand-rolled docker for servingserving.create_endpoint(...) → K8s Deployment + HPA + drift
MLflow model-versions/transition with no auditmodels.transition(...) writes hash-chained audit + approval gate
Manual drift detection scriptsserving.save_drift_baselines() + automatic PSI checks

Day-1 setup

pip install geneflow
export GENEFLOW_TRACKING_URI=https://api.genedata.io
export GENEDATA_PAT=$(genedata auth token)
export GENEFLOW_TENANT_ID=ACME

Confirm: gfctl experiments list should return without errors.

Workflow 1 — Train + log a run

import geneflow
geneflow.set_tracking_uri("https://api.genedata.io")
geneflow.set_tenant("ACME")

with geneflow.start_run(experiment_name="fraud_v3", run_name="xgb-tuned") as run:
    geneflow.log_params({"max_depth": 7, "lr": 0.05, "subsample": 0.8})
    for epoch in range(1, 11):
        geneflow.log_metric("auc",      0.7 + epoch * 0.02, step=epoch)
        geneflow.log_metric("log_loss", 0.5 - epoch * 0.03, step=epoch)
    geneflow.log_metric("cost_usd", 4.31)
    geneflow.set_tag("dataset_sha", "2026-05-12-train.parquet:sha256=...")

Verify in UI: /dashboard/geneflow/ml → experiments → click your run.

Workflow 2 — Register a model version

from geneflow import models, artifacts

artifacts.upload(run.run_id, "/tmp/model.pkl", dst_path="model/model.pkl")

mv = models.create_version(
    name="fraud-detector",
    source=f"runs:/{run.run_id}/model",
    run_id=run.run_id,
    description="XGBoost v3 — tuned",
    tags={"framework": "xgboost", "language": "python"},
)
print(mv["model_version"]["version"])  # → 3

Workflow 3 — Compare candidates ($ + metrics)

diff = models.compare("fraud-detector", a=2, b=3)
# Or in the UI: /dashboard/geneflow/ml/compare?model=fraud-detector&a=2&b=3

diff has metricsDiff and costDiff — use these to decide which version to promote.

Workflow 4 — Save a drift baseline (do this BEFORE serving)

from geneflow import serving
import numpy as np

# Compute the histogram on your training data
hist, edges = np.histogram(X_train["transaction_amount"], bins=10)

serving.save_drift_baselines(
    model_name="fraud-detector",
    model_version=3,
    features=[{
        "feature_name": "transaction_amount",
        "feature_type": "numeric",
        "histogram": {"buckets": hist.tolist(), "edges": edges.tolist()},
        "mean": float(X_train["transaction_amount"].mean()),
        "stddev": float(X_train["transaction_amount"].std()),
        "sample_size": len(X_train),
    }],
)

Workflow 5 — Deploy + monitor

ep = serving.create_endpoint(
    name="fraud-prod",
    model_name="fraud-detector",
    model_version=3,
    instance_type="cpu-large",
    min_replicas=2,
    max_replicas=10,
    drift_check_enabled=True,
    drift_psi_threshold=0.2,
)

# Poll live metrics
m = serving.get_metrics(ep["id"], window_minutes=60)
print(f"QPS={m['qps']:.2f}  p95={m['p95LatencyMs']}ms  err={m['errorRatePct']:.2f}%")

# Run a drift check (also runs automatically per `drift_check_window`)
serving.check_drift(ep["id"], window_minutes=60)
alerts = serving.list_drift_alerts(ep["id"], only_open=True)

Workflow 6 — Ship a new version (rolling)

serving.update_endpoint("fraud-prod", model_version=4, reason="precision win on holdout")
# Endpoint URL stays the same; K8s does a rolling update with maxUnavailable=0

If a Production stage transition is required:

models.transition("fraud-detector", 4, to_stage="Production", reason="canary green")
# If approval is required, an audit row records the request; an approver must call
# the approve endpoint before the transition becomes active.

Common gotchas

  • Cost not showing? You must call geneflow.log_metric("cost_usd", ...) explicitly — runs don't infer it.
  • Drift alerts never fire? Did you save a baseline first? Check gfctl drift baselines fraud-detector --version 3.
  • Endpoint stuck in CREATING? Look at gfctl endpoints show fraud-prod — the most recent revision row has the failure reason.
  • mlflow CLI still works — pointing it at GeneFlow works, but you lose the GeneFlow-native features (cost, drift). Prefer gfctl.

Where to go next