Builder lab · Module 6 · 70 min
Capstone: Ship a Bounded Healthcare GenAI Feature
Take one synthetic workflow end to end — scope, contract, grounding, tools, gates, validation, evaluation, release decision and handover — and defend every boundary.
After this module you can
Take one synthetic healthcare workflow end to end — scope, output contract, grounding, tools, gates, state, validation, evaluation, release decision and handover — and defend every boundary you set.
- State a user, a task and explicit non-goals before choosing any technology.
- Pick an output contract that matches what the workflow must guarantee.
- Say what must be grounded and in which governed source.
- Select tools, classify read and write, and place the gates.
- Define state, stop conditions and validation checks that make the run auditable.
- Turn an evaluation result into a release decision with named limitations.
Modules 1–5 each built one layer. The capstone is the assembly, and the part that is genuinely hard: deciding what the feature will not do.
Step 1
Scope one workflow
Pick the feature you will take end to end. Read the non-goals as carefully as the purpose — they are the part that keeps the build bounded.
Everything here is simulated. Every workflow, record and knowledge source below is synthetic. The end-to-end run is a deterministic function over your selections — it is a design review, not a test of a real system.
Turns a synthetic consultation record into one follow-up task and one draft patient message, both awaiting approval.
User: GP or specialty clinician immediately after a consultation.
Non-goals
- — No diagnosis, triage or treatment change
- — No action on any patient other than the one in the consultation
- — No message sent without clinician approval of the exact text
Drafts patient-facing discharge instructions grounded in a governed synthetic instruction library.
User: Ward nurse or junior doctor preparing discharge paperwork.
Non-goals
- — No medicine changes and no dosing content generated from model memory
- — No instruction issued that is not traceable to a library passage
- — Not a substitute for the discharging clinician's review
Extracts a fixed variable set into a structured table with a provenance span for every value.
User: Research coordinator extracting variables from synthetic records.
Non-goals
- — No inference of variables that are absent from the record
- — No clinical interpretation of extracted values
- — Not a replacement for double abstraction on the study's primary endpoints
Make a selection to continue. Every choice is stored in this browser only.
0 of 10 steps complete
Sources & evidence · 3 sources
This module cites public or consensus guidance, scholarly literature.
Content reviewed: September 2026. Publication dates of the individual sources are shown in each citation.
NIST AI Risk Management Framework — Generative AI Profile
Risk framing, oversight and documentation practices for generative systems.
Open sourceDECIDE-AI reporting guideline (BMJ 2022;377:e070904)
What early live clinical evaluation involves once a bounded feature leaves the workbench.
Open sourceWHO — Ethics and governance of artificial intelligence for health: LMM guidance
Governance expectations specific to large multi-modal models in health settings.
Open source