Industry Scenario — Healthcare

The team: a health system's informatics group + a research team. The goal: clinical summarization and a safe intake assistant — without PHI leaving the environment and with least-privilege access.

The journey

Build · 1 — connect your data

Connect the encrypted clinical-notes store (GCS) with strict access controls. Nothing is replicated into a SaaS pipeline.
Decision: the connection is the boundary; PHI-bearing sources are visible only to roles that need them.

Build · 2 — explore the catalog

Explore the catalog: researchers see fuller sets, other roles see de-identified views. RBAC + lineage make discovery safe.

Build · 3 — train the model

Fine-tune a summarization model on handoff-note structure and abbreviations (instruction/output JSONL) using the instruction preset.
Post-train with DPO to prefer evidence-based, concise output and refuse to invent dosing advice.
GPU smoke validates each run cheaply; the confidence cache flags "this recipe failed before."
Decision: single-GPU LoRA keeps runs small and reviewable.

Ship — your account, your rules

Hosted, BYO AWS, or self-hosted: start on k3ld hosted (zero infra), use BYO AWS, or fully self-host on AWS — covered by existing approvals and procurement.
Decision: no vendor security review; the path from idea to production is internal.

Monitor — keep it honest

Drift schedule on the summarizer's input distribution (note style changes over time).
Data capture for replay/audit of a model touching clinical text.

Export the tuned model as safetensors so a partner institution can run it inside their compliant environment — the weights, not an API, cross the boundary.

What you own at the end

A clinical summarizer + safe assistant, tuned and monitored, running hosted, BYO, or self-hosted.
Exportable weights for partner collaborations.
A short compliance review, because "no migration, processing where you choose" is the architecture.

Key docs

Connect · Fine-tune · Post-train · Ship & BYO · Export · Security