Reference

GeneFlow overview

What GeneFlow is, what's inside, and how it relates to MLflow.

GeneFlow is GeneData's ML & GeneAI lifecycle platform. It's MLflow-compatible at the wire level so existing mlflow.start_run() code works unchanged — and it adds the features MLflow doesn't have: a first-class prompt registry, real-time collaborative editing, native cost-per-run, hash-chained audit, model-serving with built-in drift detection, and a one-click migration path from existing MLflow installs.

What's inside the platform

Capability	Doc	Status
Experiments, runs, metrics, params, tags	api-reference.md	✅
Artifact upload / download (S3 or local)	api-reference.md	✅
Model registry + approval gates	api-reference.md	✅
Model serving (Deployments + HPA + drift)	api-reference.md	✅
Prompt registry + semantic search	api-reference.md	✅
Eval sets + LLM-as-judge	api-reference.md	✅
Real-time collab prompt editor (Y.js / CRDT)	architecture.md	✅
MLflow Projects compat (LOCAL / DOCKER / K8S_JOB)	api-reference.md	✅
Lineage (feature group ↔ run ↔ model version)	api-reference.md	✅
MLflow → GeneFlow importer	migration-from-mlflow.md	✅

Quickstart

pip install geneflow

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

import geneflow

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

with geneflow.start_run(experiment_name="fraud_v3") as run:
    geneflow.log_param("max_depth", 7)
    geneflow.log_metric("auc", 0.91)
    geneflow.log_metric("cost_usd", 2.34)

That's literally MLflow code with the tracking URI pointed at GeneFlow. Everything else is additive.

Where to read next

Build & operate:
- architecture.md — services, schema, message flows
- api-reference.md — every REST endpoint
- python-sdk.md — the geneflow package
- cli-reference.md — the gfctl / genedata CLI
- observability.md — Grafana, Prometheus, alerts
- security.md — pod hardening, RBAC, network policies
Migrate in:
- migration-from-mlflow.md
Role-specific how-to:
- workbooks/ — 26 role workbooks (MLE, MLOps, DS, AIE, …)
Releases:
- CHANGELOG.md

How GeneFlow fits with the rest of the platform

                ┌─────────────────────────────────────────────────┐
                │  /dashboard/geneflow/ml — control plane UI      │
                │   • experiments  • compare  • prompts (collab)  │
                │   • endpoints    • drift    • import-from-mlflow│
                └────────────────────────┬────────────────────────┘
                                         │ Hono REST
                ┌────────────────────────▼────────────────────────┐
                │  Monolith /api/2.0/mlflow/* + /api/v2.1/...     │
                │  (transparent proxy to extracted services)      │
                └─────┬──────────────────────────────────────┬────┘
                      │                                      │
        ┌─────────────▼──────────────┐         ┌────────────▼────────────┐
        │  geneflow-service :4032    │         │  jobs-service           │
        │  (tracking, registry,      │◄────────│  (K8s project runner +  │
        │  prompts, lineage,         │         │  serving deployer)      │
        │  serving, drift, importer) │         └─────────────────────────┘
        └─────────┬──────────────────┘
                  │
        ┌─────────▼──────────────────┐         ┌─────────────────────────┐
        │  websocket-gateway :4033   │         │  PostgreSQL + pgvector  │
        │  (Y.js CRDT for prompts +  │         │  (gf_* schema)          │
        │  notebook collab)          │         └─────────────────────────┘
        └────────────────────────────┘

Three positioning sentences

vs. MLflow OSS — same wire protocol, plus prompts-as-first-class, cost-per-run, hash-chained audit, real-time collab, K8s-native runners, model serving with drift detection, and a one-click migration path.
A unified lifecycle workspace — bring experiment tracking, the model registry, prompts, and serving into one GeneFlow dashboard, with self-hosted deployment options.
vs. SageMaker — model lifecycle is in one place, prompts are first-class, costs roll up per tenant in real time, and the K8s deployer is portable across clouds.

See docs/marketing/GENEFLOW-vs-MLFLOW.md for the full comparison.