Workbook

Product Manager for Data/AI Platforms

You own: what GeneFlow ships next, how it's priced, who it's for, and the metrics that prove it's working.

What's already shipped

See CHANGELOG.md. Six rounds:

  1. Schema + tracking + registry + prompts + REST + SDK + service extraction
  2. Artifacts + agent tools + dashboard UI + MLflow Projects
  3. K8s runner + compare UI + collab editor backend + tests + marketing
  4. K8s Job controller + WS gateway + prompt editor frontend + CLI + import script
  5. Runner RBAC + MLflow importer (TS-native) + WS gateway dashboard + soak test
  6. Model serving + drift monitoring (closes the lifecycle)

Marketing positioning: docs/marketing/GENEFLOW-vs-MLFLOW.md.

Roadmap candidates (next 3–6 months)

PriorityItemWhy
P0Canary / shadow traffic routing for endpointstraffic_split_pct + shadow_version columns exist; need router
P0Auto-retrain on critical driftWe have alerts but not yet the loop
P1Feature Store engineLineage refs exist; build the actual online/offline store
P1SageMaker importerAfter MLflow, second-largest migration source
P1Drift PSI per-feature thresholdsCurrently per-endpoint only
P2Model card auto-generation from gf_model_versionsCompliance ask, EU AI Act
P2Webhooks on stage transitionsCI/CD integration
P2Notebook collab (WS gateway already supports kind=notebook)Reuse the editor
P3Federated learning service (already extracted)Build out
P3Custom drift metrics (JS divergence, KS for categoricals)DQE feedback

North-star metrics

  • Adoption: tenants with > 100 runs/week
  • Retention: tenants with > 10 endpoints in Production
  • Quality: drift alerts acknowledged within 4 hours
  • Migration win: tenants who migrated FROM MLflow, signed annual

Pricing levers

  • Per-endpoint hourly (already tracked in gf_endpoints.hourly_cost_usd)
  • Per-inference markup
  • Storage tier (artifacts in S3, charged-back)
  • Premium: SOC2 audit pack, HIPAA-compliant config, multi-region

Competitive frame

CompetitorBeats us onWe beat them on
SageMakerAWS-native depth, batch transformsMulti-cloud, prompts, no Lambda-style cold starts
Vertex AIGoogle ML stack tie-inMulti-cloud, open-source-aligned
WhyLabs / ArizeDrift detection sophisticationLifecycle integration in one product
LangSmith / HeliconeLLM-only depthClassical ML + LLM in one product

Where to go next