UCOP

UCOP AI Steering Committee — research, synthesis, and analytics platform.

A Next.js web app over the AI governance baseline dataset (20 entities, 262 data points, 10 dimensions, v0.7.0) and the UCOP AI Steering Committee directory (23 members), with a grounded Claude copilot that cites every claim back to a specific entity or member. Organized around the three pillars and eight Opportunity Areas of the UCOP North Star.

Surfaces

Route	Purpose
/	Compare matrix — pick a dimension and a set of entities, read the same question answered for each.
/expertise	Committee directory with filters for expertise, opportunity area, sector, and AI relationship. List and matrix views.
/chat	Claude-powered research copilot grounded in the baseline + committee directory. Inline [entity_id] and [member-id] citations open a side drawer.
/baseline, /baseline/[id]	Per-entity detail pages over the baseline JSON.
/about	Strategy, the ten dimensions, method, and known gaps.
/admin/enrich	Review queue for the monthly enrichment run — approve or decline each proposed change and publish. Admin-gated.

Memo drafting and comparison are scaffolded but hidden from the global nav while content is finalized. The underlying routes (/memos, /memos/new, /api/memos) and admin gate are unchanged.

Status

	Step	What it adds
✓	Step 1 — Baseline Explorer	Compare matrix on /, entity detail pages, NFI design tokens. No runtime dependencies.
✓	Step 2 — Grounded Chat	Streaming Claude chat with the baseline + committee directory in a prompt-cached system prompt, inline citation chips wired to side drawers.
✓	Expertise directory	/expertise with searchable, filterable committee directory and a member × OA matrix view.
☐	Step 3 — Persistence + Shares	Postgres + Drizzle, saved chats, admin cookie gate, read-only share links.
☐	Step 4 — Memos + Comparison	Memo drafting (scaffolded; nav entry hidden for now), seeded published memos.
☐	Step 5 — Polish	Copy pass, custom domain, analytics.

Full plan: docs/v1-plan.md.

Quickstart

npm install
npm run dev

Open http://localhost:3000. No environment variables are required for Step 1 — the baseline JSON is read at build time and every page is prerendered.

Type check and production build:

npm run typecheck
npm run build

Repository layout

app/          Next.js App Router pages and API routes
components/   React components aligned to docs/seed-style-guide.md
content/      Single source of truth for pillars, OAs, research topics
lib/          Typed loader and query layer over the baseline JSON
data/         Baseline dataset (v0.7.0), committee directory, and enrichment tooling
docs/         Plan, principles, style guide

Key files:

• data/uc_ai_baseline.json — entity dataset (source of truth, v0.7.0)
• data/ucnfi-committee/ — committee directory records, schema, and enrichment strategy
• data/ENRICHMENT_LOG.md — dataset version history and gaps
• data/enrich/changesets/ — monthly proposals. {id}.md awaits review at /admin/enrich; {id}-auto.md was published automatically under policy; .rejected.json and .suppressed.json record what never made it in and why
• lib/enrich/diff.ts — the policy deciding what may publish without a human
• lib/enrich/apply.ts — the only module that writes canonical data, and the gate it enforces
• content/northstar.ts — pillars, Opportunity Areas, research topics
• lib/baseline.ts — typed accessors (listEntities, getEntity, queryBaseline, baselineStats)
• lib/committee.ts — committee directory loader and facets
• docs/v1-plan.md — approved v1 implementation plan

Design system

Design tokens in app/globals.css derive from the seed style guide and the NFI color palette — warm-neutral base, single brand accent (#005581), and a disciplined set of semantic tokens. The full mapping lives under NFI color tokens in docs/v1-plan.md.

Fonts fall back to system sans for UI and system serif for memo body. No external font loading is wired up yet — add Inter / Source Serif 4 via next/font in Step 5 if desired.

---

Deployment — Vercel

The app is a stock Next.js 15 project. Vercel's framework preset auto-detects everything; install command, build command, and output directory can stay on their defaults.

General project settings (applied once at import):

• Framework preset: Next.js (auto)
• Root directory: repo root (./)
• Node.js version: 20.x
• Function region: sfo1 or pdx1 for low-latency California access (Project Settings → Functions → Region)

Step 1 — Baseline Explorer

No configuration needed.

1. Import the GitHub repo into Vercel.
2. Accept the auto-detected Next.js preset.
3. Deploy.

Every page is prerendered at build time (28 static pages, including all 20 entity detail pages), so the deployment is effectively a static site with zero runtime dependencies. No environment variables, no database, no secrets.

Step 2 — Grounded Chat

Adds the Anthropic Claude SDK (@anthropic-ai/sdk) pointed at the UCSD TritonAI LiteLLM proxy — the sole LLM provider — and a streaming /api/chat route with prompt caching of the full baseline. The SDK is just the wire format here: /api/chat calls an open-weight model on the proxy by default (CHAT_MODEL), not Claude.

Environment variables (Project Settings → Environment Variables — add to Production, Preview, and Development):

Name	Value	Sensitive
LITELLM_API_KEY	TritonAI LiteLLM bearer token	✓
LITELLM_BASE_URL	Optional; defaults to https://tritonai-api.ucsd.edu	—

Function duration. The /api/chat route will export maxDuration = 60 so Vercel allows long-running streams. On Vercel's Hobby plan, Serverless Function duration is capped — upgrade to Pro if streaming responses are getting truncated.

Runtime. Node runtime (not Edge) so the Claude SDK's prompt caching and streaming helpers work without shimming.

No database yet — chat history lives in the browser (URL/localStorage) until Step 3.

Step 3 — Persistence + Share Links

Adds Postgres + Drizzle ORM, saved chats, admin cookie gate, and /share/[slug] read-only resolver.

1. Provision Postgres. In Vercel: Storage → Create Database → Neon Postgres. Link it to the project. Neon auto-injects DATABASE_URL (and POSTGRES_URL, POSTGRES_URL_NON_POOLING, etc.) into Production and Preview environments.

2. Add environment variables:

Name	Value	Sensitive	Notes
ADMIN_PASSWORD	Strong random string	✓	Gates Zach's editing routes via an HTTP-only cookie
NFI_BASE_URL	https://<project>.vercel.app or your custom domain	—	Used to build absolute share URLs

3. Run migrations once against the Neon database from your local machine:

# Pull env vars from Vercel into a local .env
npx vercel env pull .env.local

# Push the Drizzle schema to Neon
npm run db:push

The db:push script will be added in Step 3 and wraps drizzle-kit push. Preview deployments automatically share the same database in this setup — if you want isolated preview DBs, create a second Neon branch and scope DATABASE_URL to the Preview environment.

Step 4 — Memos + Comparison

Memos are markdown files under content/memos/ and are pre-rendered at build time. New memos can be drafted from /memos/new and are committed to the repo via the GitHub Contents API — Vercel's Git integration rebuilds the site and publishes the memo (~60s).

Environment variables:

Name	Value	Sensitive	Notes
ADMIN_PASSWORD	Strong random string	✓	Gates /memos/new and /api/memos via an HTTP-only cookie (HMAC-signed; rotating invalidates sessions)
GITHUB_TOKEN	Fine-grained PAT	✓	Scoped to z6johnson/ucnfi only, Contents: Read and write. Rotate on expiry.
GITHUB_REPO	z6johnson/ucnfi	—	Target for memo commits
GITHUB_BRANCH	main	—	Optional, defaults to main

Add these to both Production and Preview environments — a memo submitted on a preview deploy commits to main and triggers a production rebuild.

Branch protection. Don't enable "Require a pull request before merging" on main; the GitHub Contents API writes direct commits. If branch protection becomes a requirement, switch the memo route to a branch-and-PR flow.

The comparison matrix is a pure read over the baseline JSON — no additional config.

Step 5 — Polish

No required config changes. Optional:

• Custom domain — Project Settings → Domains → add your hostname and follow Vercel's DNS instructions. Update NFI_BASE_URL to match.
• Vercel Analytics — toggle on from the project dashboard for page-view counts and Web Vitals.
• Seed memos — once /api/memo/generate is live, seed a handful of published committee memos (campus maturity gradient, governance gap summary, health AI snapshot) and copy their share URLs into the committee comms.

---

Environment variable summary

Variable	Step introduced	Purpose
LITELLM_API_KEY	Step 2	TritonAI LiteLLM bearer token for chat + memo generation
LITELLM_BASE_URL	Step 2	Optional LiteLLM proxy URL (defaults to https://tritonai-api.ucsd.edu)
CHAT_MODEL	Step 2	Optional. Model ID used by /api/chat. Defaults to api-glm-5.3 (open-weight).
CLAUDE_MODEL	Step 2	Optional. Model ID used where Claude is specifically required (the Brief; web_search-tool calls). Defaults to claude-sonnet-5.
DATABASE_URL	Step 3	Neon Postgres connection string
ADMIN_PASSWORD	Step 3	Admin cookie gate for editing routes
NFI_BASE_URL	Step 3	Absolute URL used when generating share links
GITHUB_TOKEN	Step 4	Fine-grained PAT — Contents: Read and write to commit memos and enrichment decisions, Actions: Read and write to start the enrichment publish run
GITHUB_REPO	Step 4	owner/repo target for memo commits
GITHUB_BRANCH	Step 4	Optional, defaults to main

See .env.example for the local template.