Luma Stats

Dashboard for Claude ambassadors to triage Luma event registrations.

Drop your CSV, run one command, get charts, triage tools, an approved list, and an optional team-organization board.

Prerequisites

• Node.js (LTS), pnpm, Claude CLI

Quick Start

git clone <this-repo> && cd luma-stats && pnpm install
# Export your Luma guest list as CSV → data/list.csv
pnpm process    # Claude classifies and scores candidates → data/processed.json
pnpm dev        # http://localhost:3000

How It Works

pnpm process calls Claude CLI to:

• Detect Luma vs custom form columns
• Classify qualitative fields (roles, industries, interests)
• Infer synthetic dimensions: gender (from name), technicality (Technical / Non-technical / Mixed-Unclear), industry (sector from workplace)
• Compute a 0–100 relevance score
• Decide which charts to show
• Output data/processed.json

Views

• Dashboard — Charts and stats
• Triage — Approve/decline with balance tracking
• Approved — Filterable table; category badges are click-to-reassign dropdowns (overrides persist per event)
• Groups (opt-in) — Team organization board; see Groups (workshops & hackathons)

Updating Data

Re-export CSV, replace data/list.csv, then:

• pnpm process:update — only classifies new candidates; refreshes Luma fields for the rest
• pnpm process — re-design categories from scratch (form questions changed)
• pnpm reprocess — rebuild from cache without any LLM calls

Triage decisions and category overrides persist to data/triage.json / data/overrides.json via the app's API. They survive re-processing.

Groups (workshops & hackathons)

Optional 4th view for events that need teams. Hidden by default; enabled by running /groups-setup in Claude Code at the project root.

The slash command interactively:

1. Asks constraints — problems (e.g. Security / Healthcare), team size, min/max technical per team, optional secondary signal (e.g. Claude plan tier shown as a colored dot)
2. Identifies the team-mention column from your form
3. Runs pnpm groups:resolve — deterministic fuzzy name matching with union-find clustering (uses data/team-seeds-overrides.json for manual confirms/drops)
4. Runs pnpm groups:pack — packs seeds, balances technical count, validates

Produces three files in data/: groups-config.json, team-seeds.json, groups.json.

UI: drag-and-drop board, per-group lock, problem chips, swap popover, halts on zero-technical teams, CSV export. Re-pack any time via the "Re-seed" button or pnpm groups:pack — locked groups are preserved.

Deploying to Vercel

Vercel's filesystem is read-only at runtime, so deploy as a read-only snapshot — do curation locally.

1. Force-add the snapshot you want public: git add -f data/processed.json data/overrides.json data/triage.json (and data/groups-config.json data/groups.json if using Groups)
2. Commit and connect to Vercel
3. Protect access with Vercel Password Protection or your own auth — the UI shows names, emails, and LinkedIn URLs by design

PII

.gitignore covers all of data/ by default. Never force-add any of:

• data/list.csv — raw export
• data/team-seeds.json, data/team-seeds.draft.json, data/team-seeds-overrides.json — raw name mentions
• data/llm-cache.json — cached LLM responses keyed by candidate ID

Custom Form Questions

Claude analyzes each custom column and renders it as filter chips, charts, detail fields, or yes/no filters. Customize via scripts/prompt.md.