BALLHARD

A voice game where you interrupt evading politicians and CEOs at press conferences.

Why it matters

Every news cycle now opens with the same scene — a CEO or politician steps to a lectern and runs a deflection routine. Acknowledge, pivot, redirect, close. The cameras roll, the questions get asked, the dodges land. The whole choreography depends on one thing: nobody interrupts.

BALLHARD makes the interrupt the gameplay. You're Margot Renard, senior political correspondent. The principal is dodging. You cut in at the perfect moment with the cutting follow-up — or you let them get away with it. Six scandals from this week's news, seven AI principals with character-keeping under pressure, one Margot.

This was only possible because ElevenLabs Speech Engine ships a real-time barge-in primitive — when the player's voice hits the mic, an AbortSignal fires across the LLM stream and TTS playback halts within the same RTT. That's the cinematic moment competitors won't reproduce in a hackathon timeline.

Live

• Play: ballhard-web-production.up.railway.app
• Submission: ElevenHacks #10 — Speech Engine track

What's built

• Real-time WebRTC voice loop with a sub-300ms barge-in latency measured from VAD-rising-edge to audio-halt (see CLAUDE.md §12 spec; verified at ~250ms median over 5+ interrupts per session)
• Seven distinct AI principals — Scottish First Minister, American Tech CEO, Australian Chaebol Chair, etc. — each on its own provisioned Speech Engine seng_ with character-specific turn eagerness, voice, and deflection system prompt
• Margot interruption stings pre-rendered via TTS, weighted-random played on every confirmed barge-in
• Persistent verdict-card share loop — every round produces a shareable URL with a server-rendered Open Graph image that unfurls in Discord / X / iMessage

ElevenLabs surfaces stacked

Surface	Where it's used
Speech Engine	The core. Per-scenario seng_ provisioning with custom clientEvents (including vad_score for accurate T0 timing), per-path engine.attach() for routing one sidecar across 7 scenarios
TTS (Multilingual v2)	Margot's 6 interruption stings (pre-rendered), each principal's opening monologue (live, via Speech Engine firstMessage override)
Voice Library	All 7 principal voices + Margot — picked across accents, ages, and genders for distinctness
Music	Submission video bed (cinematic news thriller, low strings, urgent percussion, 90 bpm)
Voices REST API	Verifying voice IDs resolve before provisioning

Architecture

graph LR
    Browser["Browser<br/>(Next.js · WebRTC)"]
    Web["Web API<br/>(Next.js)"]
    EL["ElevenLabs Cloud<br/>Speech Engine"]
    Sidecar["Sidecar<br/>(Railway · per-path attach)"]
    LLM["Groq<br/>Llama 3.3 70B"]
    DB["Supabase<br/>verdict_cards"]

    Browser -->|"GET /api/session"| Web
    Web -->|"getWebrtcToken({agentId})"| EL
    Web -->|"{ token, sengId }"| Browser
    Browser -.->|"mic + audio<br/>conversationToken"| EL
    EL -.->|"WS dial → /ws/&lt;scenarioId&gt;"| Sidecar
    Sidecar -->|"chat.completions.create<br/>{signal, messages}"| LLM
    LLM -.->|"streamed tokens"| Sidecar
    Sidecar -.->|"session.sendResponse"| EL
    Browser -->|"POST /api/verdict"| Web
    Web -->|"insert"| DB

    classDef edge fill:#1a1a28,stroke:#3a3a4a,color:#e5e5ea
    class Browser,Web,EL,Sidecar,LLM,DB edge

Loading

The interrupt loop (the cinematic moment):

1. Player speaks while agent is mid-deflection → VAD rising-edge in browser → T0 = performance.now()
2. ElevenLabs sends a new onTranscript to the sidecar → the previous transcript's AbortSignal auto-fires
3. In-flight Groq stream throws AbortError, tokens cease → session.sendResponse returns
4. TTS playback in browser halts → onModeChange("listening") → T1 = performance.now()
5. latency = T1 − T0 → flashes red on the latency badge · Margot sting plays · scorecard ticks broken+1

Stack

Next.js (App Router), React 19, Tailwind 4, ElevenLabs JS SDK, Groq Llama 3.3 70B, Supabase (RLS-gated public verdicts), Railway (web + sidecar), Vaul (verdict sheet), LiveKit WebRTC (via ElevenLabs).

Run locally

git clone https://github.com/winsznx/ballhard.git
cd ballhard
cp .env.example .env.local
# fill in: ELEVENLABS_API_KEY, GROQ_API_KEY, NEXT_PUBLIC_SUPABASE_*, SUPABASE_SERVICE_ROLE_KEY
pnpm install
pnpm provision:engine --all --ws-url=wss://your-sidecar/ws   # one-time per environment
pnpm dev:engine    # sidecar on :8787
pnpm dev           # web on :3000

The sidecar must be reachable from ElevenLabs cloud (publicly resolvable WS) — Railway provides this. For local dev against a hosted sidecar, set the seng_s' wsUrl to your sidecar's public Railway URL.

Credits

Built for ElevenLabs Hack #10 (Speech Engine), May 2026.

License

MIT