HiveMind — group ideation, fused by AI

A group-ideation tool for hackathons that collects every teammate's idea privately, then fuses them. Each person submits one pitch without seeing anyone else's — openly or anonymously, their choice — and when the last one lands, AI takes the single most distinctive element out of every submission and builds four new project ideas that each carry a piece of everybody's thinking. Then the team votes, and turns the winner into a build plan.

▶ Live: hivemind.rianfernando.com · Guided demo · Architecture · Privacy model

Why it's different

Group brainstorms are decided by whoever speaks first and loudest. Every quiet person's idea evaporates, and the team converges on the first plausible thing said out loud. HiveMind removes the ordering problem entirely: pitches are collected in parallel and in private, and fusion is mechanical — every participant's idea is required to appear in every generated concept. The results show which element came from whom, so nobody has to argue that they contributed.

The part that makes people actually use it is the privacy toggle. Each person independently decides whether their name and their pitch text are public. A fully anonymous pitch still shapes all four results — it just never appears in the credits. That is enforced on the server, not in the UI: the browser has no read access to pitches at all, and the AI is handed placeholder labels instead of hidden participants' names.

And it deliberately stops short of deciding for you. The AI produces raw material; the team argues, votes, and commits. Nothing auto-advances past the reveal.

Architecture

flowchart LR
  subgraph Clients["Participants · phones and laptops"]
    H["Host<br/>creates room"]
    P["Pitchers<br/>one idea each"]
  end

  subgraph App["Next.js 15 · Vercel"]
    API["API routes<br/>service-role only"]
    MASK["Privacy masking"]
  end

  subgraph Data["Supabase"]
    DB[("Postgres<br/>rooms · ideas · votes")]
    RT["Realtime<br/>rooms row only"]
  end

  subgraph AI["AI · free tiers"]
    GEM["Gemini<br/>primary"]
    GRQ["Groq · Llama 3.3<br/>fallback"]
  end

  H --> API
  P --> API
  API -->|writes| DB
  DB --> RT
  RT -->|"content-free ping"| Clients
  Clients -->|"masked view"| MASK
  API -->|"placeholder labels"| GEM
  GEM -.->|"rate limit / bad JSON"| GRQ
  GEM --> MASK
  GRQ --> MASK
  MASK -->|"credits per privacy choice"| DB

Loading

The browser can read exactly one table (rooms, which holds nothing sensitive) and write nothing. Every mutation goes through an API route holding the service-role key. Realtime publishes only the rooms row, so pitch text is never broadcast — writes bump updated_at as a content-free "something changed" ping and clients refetch a masked view. Full detail in docs/architecture.md.

What it does

• Rooms in one step — name the event, set the group size (2–50), optionally set a pitch countdown. Share a link, a QR code, or a six-letter code.
• Private pitching — one idea per person, submitted without seeing the others. Two independent toggles hide your name, your idea text, or both.
• Live room — visible pitches stream in with 🔥💡😂 reactions and a progress bar; a countdown, when set, fires fusion automatically at zero.
• Fusion — one AI call per room extracts the most distinctive element of each pitch and combines them into four hackathon-scoped concepts, credited according to each person's privacy choice.
• Voting — one changeable vote per device, live tallies, leading idea crowned.
• Build plans — per idea, on demand: MVP features, a stack with reasoning, a role for each teammate, stretch goals, and a first-hour checklist. Cached room-wide, so only the first click pays for it.
• Presenter view — /room/CODE/present for a projector: a large QR code and live counter, then oversized result cards with live tallies.
• Export — copy or download the whole session as Markdown, votes and build plans included.

The privacy model

Hide name	Hide idea	In the live room	In the results
—	—	Maya — food-waste map…	Maya · real-time maps
✅	—	Anonymous — food-waste map…	Anonymous · real-time maps
—	✅	Maya — 🔒 pitch kept private	Maya · secret ingredient 🤫
✅	✅	Anonymous — 🔒 pitch kept private	no credit at all

Enforced at four points: no browser read access to the ideas table, a masking route for every read, placeholder labels in the AI prompt, and masking of the AI's output before it is stored anywhere browser-readable. The limits of the model are written down honestly in docs/privacy-model.md — including the fact that small rooms leak by arithmetic and that the operator can always read raw pitches.

Stack — free tiers only

Layer	Choice	Free tier
Frontend + API	Next.js 15 (App Router), React 19 on Vercel	Hobby, free
Database + realtime	Supabase Postgres	500 MB, Realtime included
AI — primary	Google Gemini	free API tier, no card
AI — fallback	Groq (Llama 3.3 70B)	free API tier, no card
3D landing	three.js + React Three Fiber	open source, renders on-device
QR codes	qrcode.react	generated client-side, no service
Feedback	Feedex	one script tag, optional

Nothing here bills. Fusion runs once per room and a build plan once per idea, so a busy event costs a handful of AI calls rather than one per person. Attribution for every third-party service is in NOTICE.md.

Run it

Two free accounts (Supabase, plus Google AI Studio and/or Groq for keys), about ten minutes.

git clone https://github.com/Rian-Fernando/HiveMind.git
cd HiveMind
npm install
cp .env.example .env.local     # fill in the five values
npm run dev

1 — Supabase. Create a free project, open SQL Editor → New query, paste supabase/schema.sql and run it. Copy the project URL, the anon key and the service_role key from Project Settings → API. (Upgrading an older deployment? Run supabase/migration-002-features.sql instead.)

2 — AI keys. aistudio.google.com/apikey for Gemini, console.groq.com/keys for Groq. Either one alone works; with both, Groq covers Gemini's rate limit.

3 — Run. npm run dev, then open the room link in a second browser or an incognito window to play a teammate.

Deploy: import the repo at vercel.com/new, add the same five environment variables, deploy. Set your custom domain as the primary so the canonical URL and the *.vercel.app redirect line up.

Feedback (optional). Setting NEXT_PUBLIC_FEEDEX_KEY to a project key from Feedex puts a feedback button on every page except the presenter view, with browser and page context attached to each report. Leave it unset — the default — and nothing is rendered or requested.

npm run dev        # local dev server
npm run build      # production build
npm run typecheck  # tsc --noEmit — same check CI runs

Project structure

app/
  page.tsx                    landing — server-rendered story + FAQ, 3D backdrop
  demo/page.tsx               guided walkthrough on sample data
  llms.txt/route.ts           machine-readable summary for AI answer engines
  room/[code]/page.tsx        the room: pitch + privacy, countdown, live feed
  room/[code]/present/        presenter view for projectors
  api/rooms/route.ts          POST — create room (code + hashed host key)
  api/ideas/route.ts          POST — submit an idea (validation + privacy flags)
  api/progress/route.ts       GET  — privacy-masked pitches + reaction tallies
  api/reactions/route.ts      POST — toggle 🔥💡😂 on a visible pitch
  api/votes/route.ts          GET/POST — live tallies, one vote per device
  api/generate/route.ts       POST — race-safe fusion, Gemini → Groq, masked
  api/deepdive/route.ts       POST — build plan per idea, cached room-wide
components/
  ResultsView.tsx             results, voting, build-plan modal, export, confetti
  CreateRoomPanel.tsx         the landing page's one interactive island
  DemoWalkthrough.tsx         the demo stages
  three/HiveScene.tsx         the scroll-driven WebGL scene
  three/HiveBackdrop.tsx      mounts it — WebGL, reduced-motion and perf guards
lib/
  ai.ts                       prompts, both providers, fallback
  device.ts                   anonymous device key + room history
  supabaseAdmin.ts            service-role client (API routes only)
  supabaseBrowser.ts          anon client (reads rooms only)
supabase/
  schema.sql                  full schema for a fresh project
  migration-002-features.sql  upgrade path for earlier deployments
docs/
  architecture.md             trust boundaries, exactly-once generation, AI layer
  privacy-model.md            the four modes, enforcement points, and the limits

Discoverability (SEO + GEO)

The landing page is fully server-rendered semantic HTML — the 3D scene is a decorative, aria-hidden canvas behind it, dynamically imported so first load stays around 110 kB. Search engines and AI answer engines read real content, not an empty shell.

• sitemap.xml, robots.txt, canonical URLs, and a permanent redirect from the raw *.vercel.app host
• /llms.txt following the llms.txt convention, advertised in <head>
• Fourteen AI crawlers allowed by name — GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, CCBot and others
• SoftwareApplication, FAQPage and HowTo structured data
• One <h1> per page, real landmarks, a 1200×630 PNG social image, and prefers-reduced-motion honoured across both CSS and WebGL

License

MIT — see LICENSE. Third-party attribution in NOTICE.md; release history in CHANGELOG.md.

Built by Rian Fernando.