Talk to sales
Workbook

Data Scientist

You own: experimentation, baseline modeling, statistical rigor, and handing off candidates to MLE for production.

What's new for you

WasNow in GeneFlow
Local notebooks with scattered metric filesmlflow.start_run() → fully captured in GeneFlow
Re-running expensive comparisons by handmodels.compare(a, b) returns metricsDiff + costDiff
Cost was opaquecost_usd is a first-class metric
Lineage was tribal knowledgelineage.upstream(run_id, kind="dataset", id="…")

Day-1 setup

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

Workflow 1 — Track a run from a notebook

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

with geneflow.start_run(experiment_name="fraud_eda", run_name="baseline-lr"):
    # Train as you do
    geneflow.log_params({"C": 1.0, "solver": "lbfgs"})
    geneflow.log_metric("auc",      0.81)
    geneflow.log_metric("log_loss", 0.34)
    geneflow.log_metric("cost_usd", 0.04)

Existing MLflow notebooks: change import mlflow to import geneflow as mlflow and you're done. Or just point MLFLOW_TRACKING_URI.

Workflow 2 — Declare data lineage

geneflow.lineage.upstream(
    run_id=run.run_id,
    kind="dataset",
    id="s3://genedata-acme/datasets/fraud/2026-05-12.parquet",
)
geneflow.lineage.upstream(
    run_id=run.run_id,
    kind="feature_group",
    id="user_features_v2",
)

Now anyone can ask: "what models were trained on user_features_v2?" → lineage.get_downstream(kind="feature_group", id="user_features_v2").

Workflow 3 — Run a sweep

GeneFlow doesn't have a built-in sweep orchestrator, but parent/child runs are first-class:

with geneflow.start_run(experiment_name="fraud_sweep", run_name="parent") as parent:
    for lr in [0.001, 0.01, 0.1]:
        with geneflow.start_run(experiment_name="fraud_sweep",
                                 run_name=f"lr={lr}", parent_run_id=parent.run_id):
            geneflow.log_param("lr", lr)
            geneflow.log_metric("auc", train(lr))

The compare UI knows about parent/child relationships.

Workflow 4 — Compare candidates side-by-side

UI: /dashboard/geneflow/ml/compare?run_a=run_abc&run_b=run_def

Shows params, all metric series, and cost_usd deltas in one view. Great for screen-sharing with stakeholders.

Workflow 5 — Hand off to MLE

  1. Promote your best run to a registered model version:

``python from geneflow import models models.register("fraud-detector") models.create_version("fraud-detector", source=f"runs:/{best_run.run_id}/model", run_id=best_run.run_id, description="DS baseline candidate") ``

  1. Add a tag on the version with the eval summary
  2. Open a JIRA / Slack to the MLE owner referencing the version number

Workflow 6 — Reproduce a teammate's run

import json
run = geneflow.get_run("run_abc123")
params = run["data"]["params"]      # set-once
tags   = run["data"]["tags"]
upstream = geneflow.lineage.get_upstream("run_abc123")

# Pull the dataset version from upstream and retrain

Common gotchas

  • Forgetting cost_usd — it's not auto-computed; you have to log it. Bake it into your training harness so it's never missed.
  • Don't log secrets in params — params are visible to anyone with read access on the experiment. Use tags/links instead.
  • Step ordering — step= is what the time-series chart uses; epochs work fine, but mixing step semantics across runs makes comparison weird.

Where to go next