AI Accelerator · Knowledge systems

Ayin Wiki

Org-wide, LLM-curated knowledge base: upload raw documents, run a fail-closed PII/PHI gate, regenerate structured wiki pages with Bedrock—and serve the same content to humans (SPA) and agents (MCP).

AWS BedrockMCPRAG / wikiPHI gateVue 3 · Quasar
Ayin Wiki home with category cards and workspace shell
Workspace home — browse categories, search, Ask AI, and admin tools in one shell

Problem

Institutional knowledge lived in PDFs, exports, transcripts, and scattered docs. Humans could not search a single source of truth; AI agents had no structured, permissioned way to pull org context. Any LLM pipeline also had to respect healthcare PHI constraints—false positives from generic PII scanners would block legitimate technical content.

Role

Principal Product Designer on the AI Accelerator Team—end-to-end product design for the Quasar SPA (reader, creator, admin), ingest/regen mental models, and agent access via MCP. Designed the pairing of wiki content with an LLM-ready Figma design system so new AI-assisted projects start with tokens, components, and patterns aligned to internal and external product UI.

Product model

  • Readers — browse, search, cross-link structured MD pages.
  • Creators — upload raw docs, choose ingest mode, trigger regeneration.
  • Admins — prompts, models, categories, quality, workspaces, members.
  • Agents — same pages via MCP (wiki.search, wiki.getPage, wiki.listCategory).

Key design decisions

Ingest modes

Rewrite for unstructured material (planner → writer fan-out). Format-only for already-authored markdown—body echoed verbatim; only frontmatter, wiki-links, and related/sources appendix generated—so well-written source is not degraded by cheaper models.

Two-stage PII / PHI gate

Comprehend DetectPiiEntities first; LLM adjudicator only when flagged. Quarantine requires both stages to agree—cutting false positives (URLs, technical names, version strings) while holding the line on real PHI. Fail closed; adjudication records persisted for audit.

Humans and agents, one corpus

SPA and MCP read the same S3 MD pages. Design system packages for LLMs mean AI-built tools inherit product-consistent UI primitives from day one.

Generated wiki page (PHI Handling Policy) with body content and feedback
Structured wiki page — generated MD with breadcrumb, model metadata, and accuracy feedback
Admin pages table listing generated wiki pages by category and model
Admin pages — category tree, model lineage, and approval controls across the corpus
Ask AI grounded Q&A surface for the wiki
Ask AI — grounded Q&A over the same pages agents consume via MCP

Architecture (summary)

Quasar SPA (S3 + CDN) → API Gateway / Lambdas → S3 raw/extracted → SQS debounce → regen worker (Bedrock) → pages + DynamoDB + CloudFront. Cognito for auth. MCP server over HTTP+SSE for agent consumers.

Outcomes

  • Unified knowledge path from raw upload to wiki page and agent-readable MCP surface.
  • PHI-aware ingest that balances safety with developer/technical content usability.
  • Design-system-for-LLM bridge so Accelerator projects stay on-brand across internal and external products.

Stack

Vue 3 · Quasar 2 · TypeScript · AWS CDK · Cognito · API Gateway · Lambda · S3 · SQS · Bedrock · Comprehend · MCP · CloudFront

← All case studies