Workbook

Solutions Architect

You own: customer-facing design — propose, scope, and sequence GeneFlow rollouts inside customer environments. You map their existing stack to ours.

What's new for you

Round 6 closed the lifecycle: train → register → serve → monitor. You can now sell GeneFlow as a complete MLflow + SageMaker + LangChain + LangSmith replacement, not just tracking.
The MLflow importer (Round 5) is the conversation-opener for "we already have MLflow."
The K8s deployer is portable across EKS / GKE / AKS — no cloud lock-in.

Discovery questions

Existing ML stack: MLflow OSS / hosted MLflow / SageMaker / homegrown / nothing?
Existing GeneAI stack: LangChain / LangSmith / Helicone / no observability?
Cloud + K8s: AWS+EKS / GCP+GKE / Azure+AKS / on-prem?
Compliance: SOC2 / HIPAA / GDPR / FedRAMP?
Tenants: single-team, multi-team-in-one-org, or true multi-tenant SaaS?
Volumes: runs/day, inferences/sec, prompts deployed?
Skill profile: who'll use GeneFlow on their side? Map to our role workbooks

Stack mapping cheat sheet

Their tool	GeneFlow equivalent
MLflow OSS / hosted MLflow	Validate tracking API and artifact access, then test the importer
SageMaker Training	MLflow Projects via K8S_JOB
SageMaker Model Registry	`gf_registered_models`
SageMaker Endpoints	`gf_endpoints`
LangSmith / Helicone	Tracking (`cost_usd` + `tokens_used`) + audit
LangChain prompt strings in code	Prompt registry + versions + stages
WhyLabs / Arize for drift	Built-in PSI w/ baselines
dbt for metric definitions	Keep dbt — wire GeneFlow gf_* as a source
AirByte / Fivetran into warehouse	Same; we publish gf_* via CDC

Reference architectures

Pattern A — Single-cloud customer (AWS)

EKS cluster with our Helm chart
IRSA-enabled geneflow-runner SA → IAM role for S3 artifact access
Aurora Postgres + pgvector
S3 artifact bucket per tenant with KMS CMK
External Secrets Operator → AWS Secrets Manager
Their CI: Bitbucket / GHA — points at GeneFlow on every commit

Pattern B — Multi-region (regulatory)

One GeneFlow stack per residency region
Cross-region audit-log replication (read-only)
Tenant data_residency_region enforced at the proxy

Pattern C — Hybrid (training on-prem, serving in cloud)

Project runner on-prem K8s
Serving on EKS via separate Helm install
Artifact bucket in the cloud, mirrored to on-prem for hot training data

Sequencing a deal

Quarter	Deliverable
Q1	Helm install + tracking migration via importer. DS/MLE point clients at GeneFlow.
Q2	Prompt registry + collab editor for AIE team. Eval sets for top 5 prompts.
Q3	Model serving for 1–2 production models. Drift baselines.
Q4	Full lifecycle: project runners, all prompts in registry, all models served, drift alerts on-call.

Common gotchas

Don't underestimate the prompt-registry sell — most customers don't realize they NEED one until you show them how often prompts change silently.
Don't sell drift as "AI Observability" — it's PSI, not magic. Be honest about what it detects vs. doesn't.
MLflow drop-in is the foot-in-door — they can adopt without committing.

Where to go next

docs/marketing/GENEFLOW-vs-MLFLOW.md — the comparison
architecture.md — for technical proposals
18-cloud-data-architect.md — for cloud-architecture-specific work