memories/ai-portal-rollout-status-41a28798.md

memory

The AI Portal platform transition is moving past active build phases. Current focus is on rollout, stakeholder engagement, and initial content approval processes. Early engagement with content approvers on the revenue side is underway.

Status Update (Aug 24-26, 2026)

  • Project Rename: Formally changed from "AI Hub" to "AI Portal" (APMT-223).
  • Launch Window: Hard launch target is Sept 16, 2026, no later than end of month.
  • Governance: Implemented hub-and-spoke. [[John Kipe]] (Revenue) and [[Bridget Harvey]] (Marketing) are domain curators. Default asset state is hidden; admins must publish.
  • Roadmap Highlights: Developing an onboarding flow with default asset pre-loading per role; introducing dependent dropdowns for audience targeting (Dept -> Team) to replace fragile tags.
  • Adjacent Initiatives: GTM "Brain"/context layer being explored; team prefers internal build-on-top-of-Portal rather than parallel third-party tooling.
  • Recent Jira Milestones: APMT-184 In Review; SCIM/SSO foundations (APMT-204/218/268/320/322) largely complete.

Detailed Current State (as of Aug 26, 2026)

  • Technical Status: Live but unannounced; accessible via Okta SSO with automated onboarding. Admin-gated publishing model active.
  • Major Blocker: Agent curation. Estimated 700-1,200 agents in [[Glean]] require review.
  • Open Issues:
    • Onboarding: Selection cap bug and "select all" failure.
    • Targeting: Moving from fragile tags to dependent dropdowns (Dept -> Team) sourced from [[Docebo Board]].
    • Provisioning: JIT regression affecting some users (e.g., John Kipe).
    • Integration: OAuth token flow (Okta-to-Glean) to distinguish "Open" vs "Request Access" agent states.
  • Governance: Hub-and-spoke model confirmed. [[John Kipe]] (Revenue) and [[Bridget Harvey]] (Marketing) identified as domain curators.
  • Ownership Summary: Overall project: [[Samuel Adams]]; Tech Build: [[Giuseppe Bonanno]]; Features/Personalization: [[Preet Patel]]; Revenue/Marketing Curation: [[John Kipe]]/[[Bridget Harvey]].

Status Update (Aug 28, 2026)

  • MVP Status: Officially MVP-complete as of 2026-08-28 (search, Okta-gated access, contribution flow, admin checks).
  • Open Follow-ups (Feedback):
    • Team Targeting: Sub-audience field implementation is blocked pending an Okta SCIM department-field fix (Jira APMT-330).
    • Asset Types: Potential expansion to searchable "skills"/guides is currently on hold as possible scope creep.
    • Auth Workaround: Implemented a Glean OIDC/bearer-token workaround for the "request access" flow to mitigate the lack of SAML IdP support (requires 7-day token refresh).

Phase 2 Scoping (Aug 31, 2026)

Team scoped a "solutioning" agent for the AI Portal:

  • Overlap Detection: Automated analysis for duplicate/refinement detection in the skills repo, including build approach and next-step recommendations (citizen-dev request vs. DIY).
    • Owner: [[Preet Patel]] (responsible for synthetic/real test case generation).
    • Discovery: Required with power users [[Bridget Harvey]], [[Amanda]], and [[John Kipe]].
  • In-Portal Skill Viewer: [[Giuseppe Bonanno]] is building a markdown viewer for skill content, enabling non-GitLab users to access skill documentation directly within the portal.

Portal Documentation & Search Strategy (as of Aug 2026)

  • Guides Architecture: Current guides (vscode-setup, cowork-setup, etc.) are static markdown files (/src/lib/guides.ts and public/guides/*.md), rendered client-side without DB persistence or search indexing.
  • Search Gap: Portal search is limited to asset metadata via Postgres plainto_tsquery. Guides are currently invisible to search.
  • Discovery Plans: Documented as 'M5' in docs/strategic-review/02-plan-vs-codebase-misalignments.md. External enterprise discovery (Glean indexing of guide pages) is tracked as a separate workstream (M6) discussed with [[Sam Irizarry]] on 2026-08-19.
  • Scope Note: Phase 2 (APMT-332) focuses on the Solutioning Triage Form; guide/search infrastructure is currently out-of-scope for the active Phase 2 project plan.

Architectural Shift: Asset Defaults (Sep 2, 2026)

  • Decision: Department/team-based asset defaults are now keyed off the [[Glean]] sync job (which already provides asset owner emails), bypassing reliance on Okta SCIM department data due to ongoing SCIM inconsistency (APMT-345).
  • Rationale: Okta SCIM data is incomplete and Docebo Board lacks sufficient themes. The existing Glean sync job provides a reliable alternative source.
  • Status: This supersedes the previous dependency on the Okta SCIM department-field fix (APMT-330).
  • Future Consideration: The team is maintaining the GET /api/admin/users extension for department filtering, as it remains useful for testing.

Phase 2 Update (Sep 9, 2026)

  • Agent Migration: Shifted from locally-run Qwen to [[AWS Bedrock]] + Cohere.
  • Local-Dev Model Standard: Per guidance from [[Samuel Adams]], developers must build against Bedrock via AWS SSO rather than local models to ensure behavioral consistency between dev and prod environments.
  • Financials: Spend is negligible unless running heavy Opus volume; flag high-volume usage to AI Operations Team.