HackTX 2026, the largest student-run hackathon in Texas (UT Austin, October 24–25, 2026)
Bring any song to the enchanted ball. Dance Fantasy listens to a song you choose, choreographs a dance routine for it that has never existed before, and teaches it to you with a 3D storybook dancer while your webcam scores every beat. Dance alone, with friends in front of one laptop, or in an online ballroom with friends on their own phones.
Play at DanceFantasy.tech (also at midnight-ball.vercel.app) · Once upon a time…
| Track | What it does in Dance Fantasy | |
|---|---|---|
| 🎺 | Meta Muse | Writes the Royal Herald's live intro and final verdict from your name, song, score and best move |
| ✨ | Google Gemini (MLH) | The herald's expressive voice, turns a viral dance into playable moves, backup writer |
| 🔎 | SerpApi | Finds the song's viral dance online; those moves get pinned to the chorus and score 2x |
| 🐯 | Tiger Data (MLH) | Hypertables + continuous aggregates: ballrooms, friends, history, "hardest moments", Hall of Fame |
| ☁️ | Vultr (MLH) | An always-warm song-analysis worker on Vultr Cloud Compute, so uploads skip serverless cold starts |
| 🌐 | .tech domain (MLH) | The game lives at DanceFantasy.tech |
| 🤖 | Cursor · Grok | AI coding, automations and workflows we used to build the game |
Dance games only let you dance to the songs someone already choreographed. We wanted the opposite: any song, even one you recorded yesterday, becomes a playable dance. Then we wrapped it in the fairy tale of the Twelve Dancing Princesses, who slip out every night to dance until their slippers wear through, because a dance game should feel like a night at the ball.
- Any song becomes an original routine. Upload a track; the server finds its beats, bars and sections (verse, chorus, bridge…) and choreographs moves that land on the downbeats, repeat on every chorus, and match the difficulty you pick.
- Moves from real dancers. Pick Classic (our 87 storybook moves), Real dancers (249 clips cut from motion-captured dancers across 10 genres, from AIST++) or Mix.
- A 3D princess or prince teaches it. They dance the routine in a midnight ballroom, with the next moves sliding in as 3D cue cards.
- Your webcam scores every beat: Perfect / Good / OK / Miss, combos, stars, your best and hardest moves, and a highlight reel of your best streaks. Pose tracking runs entirely in your browser, so no video ever leaves your device.
- The song's famous dance is built in. If a song has a viral dance, the game finds it online and puts those moves into your routine, worth double points.
- A Royal Herald cheers you on by name, out loud, as your combo climbs.
- Dance together, three ways:
- Ballrooms: friends join from their own phones by QR code, chat, and start on the same beat. Play Versus (who dances best) or Crew (bonus points when everyone hits a beat together).
- Party mode: 2–4 friends in front of one laptop, each tracked and scored separately.
- Dance cards: an automatic guest profile (sign in with a username and password to keep it), friend codes, invites, recent dance partners and history.
- Learn mode loops any 8 counts at half or three-quarter speed, with tips like "left arm late".
- A Hall of Fame and badges. Every song has its own leaderboard (all time, this week, today, or just you and your friends) with a podium for the top three, and the results screen tells you where you placed. Your dance card collects 16 fairy-tale badges, from First Dance to Stroke of Midnight, with a progress bar on each one you haven't earned yet.
| Title menu | Online ballroom |
|---|---|
 |
 |
- Truly any song. Other dance games ship hand-made routines for a fixed song list. We generate an original, playable routine from the audio itself in seconds: beat tracking, section analysis, then a search over 87 moves that keeps the dance musical (chorus moves repeat, big moves land on big beats).
- 3D from 2D. Our routines are 2D poses, so the 3D dancer recovers depth from how short each limb looks on screen and the body's turn from the width of the shoulders. No 3D models, no motion capture: the dancer is built from simple shapes in three.js.
- Multiplayer that runs on serverless. Ballrooms work with no game server: devices sync to the server clock and start from one shared timestamp, so phones and laptops dance the same beat.
- A storybook you can play. Every screen is a hand-built fairy-tale sticker book: room codes in a magic mirror, a clock that strikes twelve when the song ends, and party awards like the Glass Slipper.
Muse Spark writes the herald's big moments live, from what you did: it throws open the ballroom doors with your name and the song, and delivers the final verdict with your score, grade, best move and longest combo. During the song the herald calls out every combo and cheers you on with recorded lines, so it never lags behind your feet. Muse writes the intro while you calibrate, so it's ready the moment the music starts. Gemini is the fallback, and a hand-written line covers anything neither answers in time, so the show never waits.
- The herald's voice. Gemini text-to-speech performs every line with a directed delivery ("an ecstatic royal herald announcing the star of the ball"). Lines are voiced once and cached, and the music dips while the herald speaks.
- Turning a viral dance into choreography. Gemini reads SerpApi's search results, names the song's signature dance, and maps it onto moves in our library so the dance becomes playable.
- Backup writer for the herald when Muse is busy.
When you pick a song, SerpApi searches Google and YouTube for its dance challenge. The result doesn't just show a link: Gemini identifies the moves, our choreographer pins them onto the chorus, an in-game banner announces "Signature: YMCA arms", and those beats score double. The source video is linked so you can watch the original.
Tiger Data (Postgres + TimescaleDB) holds everything multiplayer and social: ballrooms, dancers, chat, accounts, friends and invites. Every round and every beat danced in a ballroom is stored in hypertables, and a continuous aggregate rolls those beats up into each song's "hardest moments", the beats that trip up dancers across every ballroom. A second continuous aggregate adds up each dancer's score per day for the Hall of Fame's top dancers of the week. It also keeps the paid AI and search calls within daily and per-visitor limits.
Analyzing a new song is the one heavy job in the game: librosa's beat tracker compiles itself to machine
code on first use, which costs a fresh serverless instance 20-30 seconds. We built a Vultr Cloud
Compute worker for it: the same FastAPI app on an instance in Dallas that stays warm, so an upload is
analyzed in about a second.
deploy/vultr/provision.sh creates the instance through the Vultr API, and cloud-init.yaml installs the
server, the numba cache and Caddy for HTTPS. With ANALYSIS_WORKER_URL set, the Vercel API hands each
uploaded song to the worker (server/app/vultr_worker.py); the worker reads it from the shared Blob store
and caches the analysis where every instance sees it.
The ball has its own address: DanceFantasy.tech (and www.), served over
HTTPS by our Vercel deployment.
We used Cursor for AI-assisted coding, and Grok automations and workflows during development, alongside Claude Code's parallel agents that built and reviewed features on separate branches.
MediaPipe (Google) for in-browser pose tracking · librosa for beat and structure analysis · three.js for the 3D dancers · React + TypeScript + Vite · FastAPI (Python) · Vercel (web, API and Blob storage).
- Listen. librosa finds the tempo, beats, downbeats and song sections, grouping repeats so every chorus is recognized as the same music.
- Choreograph. A beam search picks moves from the library to fit each section, snapped to the downbeats, with signature moves pinned onto the chorus. The same song and settings always make the same routine, and "New routine" rolls a fresh one.
- Dance. MediaPipe tracks 17 body joints from your webcam in the browser. A quick clap-along calibration measures your camera's delay, then each beat compares the direction of your arms, legs and torso with the dancer's.
- Together. Ballroom devices sync to the server clock and post their beat scores about once a second, which powers the live leaderboard, Versus beats won and Crew sync streaks.
- Depth from flat poses. Moves are stored as 2D joint positions, so making a convincing 3D dancer meant reconstructing depth (knee lifts, spins) from limb foreshortening.
- Webcam lag. Cameras show you about 100–300 ms late, which made perfect dancers score "late". The magic-mirror calibration measures that delay with a clap-along and subtracts it.
- Audio analysis on serverless. librosa and its numerical stack are about 430 MB, well past normal serverless limits, so the API runs as a large function with a shared cache.
- Real-time play without a game server. We built rooms on short polling and a shared start time, so up to 8 phones and laptops dance the same beat on a serverless backend.
DanceFantasy.tech (or midnight-ball.vercel.app) in Chrome with a webcam. Pick a song, press Start the ball, stand back so your whole body is in the mirror, and dance. No webcam? Turn on Autoplay in the lobby to watch the bot dance.
To run it locally (Node 22+ and uv):
make setup # Python, npm packages and the pose-tracking model
make dev # opens the API on :8000 and the game on http://localhost:5173API keys are optional (see .env.example): without them the herald uses hand-written lines and ballrooms use a local database. Technical docs: shared/README.md.
Toyeshh Medikonda, Luv Udeshi, Suchay Kommisetty, and Gatik Goyal.
- Claude Code (Anthropic) helped us scaffold and write the code, including parallel agents that built and reviewed features on separate branches.
- Cursor for AI-assisted coding alongside Claude Code.
- Grok automations and workflows during development.
- Meta Muse, Google Gemini and SerpApi power features in the game, as described above.
- MediaPipe (Google) for pose tracking.
Real-dancer moves: AIST++ annotations (© Google LLC, CC BY 4.0), derived from the AIST Dance Video Database (research and non-commercial use). We ship only our processed 2D clips, never AIST video or music.
Sample songs: "Vibe Ace" by Kevin MacLeod and "Let's Go Fishin'" by Karissa Hobbs (both CC BY 3.0, from Free Music Archive via librosa's example data), plus synth tracks we generated (CC0). Fonts: Fraunces and Figtree. Webcam frames and poses never leave your browser.