Kinneas

Singapore Airlines — Concept Site A reimagining of the Singapore Airlines homepage as long-form cinema instead of a fare-finder utility. Full-bleed plane footage, parallax destination strips, Instrument Serif typography, and "liquid glass" surfaces — with the full booking, KrisFlyer, and check-in IA preserved underneath the editorial frame. Built with React 19, Vite, TypeScript, and Tailwind. Static, no backend, deploys to any CDN. ElevenLabs powers two voice surfaces: Live Concierge — A real-time conversational agent (floating button, bottom-right) using @elevenlabs/react. Opens a WebSocket session to an ElevenLabs Conversational AI agent, streams the user's mic up, plays the agent's voice back, and surfaces a live transcript. Users can ask about destinations, fares, baggage, or KrisFlyer in their own voice. Ambient destination narration — Each parallax strip (Bali, Tokyo, Paris, NYC, Sydney, London, Zurich) auto-plays a short cinematic narration when scrolled into view. Generated at build time via the ElevenLabs TTS REST API with tuned voice settings for an editorial delivery, then shipped as static .mp3s — so runtime never hits the API. The split is deliberate: TTS is precomputed and free at runtime, while the live agent is opt-in only when a user engages — keeping the page fast by default.

Submitted 7 May 2026

Voice Thief is a noir heist puzzle game where the player has no voice — but everyone else does. The core mechanic: record NPCs in the wild, steal their voices, and use them to manipulate voice-locked doors, intercoms, and phone calls to break into a bank vault before the clock runs out. It solves a creative design problem — how do you make voice AI the gameplay itself, not just a UI layer? Every locked door in the game is a voice puzzle. The vault accepts only the manager's calm voiceprint. The hallway intercom belongs to the secretary. Use a panicked recording and you trip the alarm. The player must social-engineer the right people into saying the right things, in the right emotional state, at the right time. ElevenLabs powers the entire voice layer: all 14 NPC dialogue clips are rendered with ElevenLabs TTS to give each character a distinct, human voice. The phone call system uses ElevenLabs to synthesize real-time impersonation audio when the player makes a call in a stolen voice. The architecture is IVC-ready for live Instant Voice Cloning flows on auth surfaces. Zed was used as the primary editor throughout the build.

Voice Pictionary with a live AI partner AI Pictionary is a timed drawing game where your sketch is the clue and an AI is the guesser. It keeps firing spoken guesses as you draw, so the experience feels like a real back and forth instead of a single static answer. Why it exists: People scroll past “cool tech” unless the payoff is instant. Here the payoff is obvious in a few seconds: you draw, the AI talks, and the round ends with a voiced reaction when you win or lose. ElevenLabs: All guess lines and end of round reactions use text to speech, so audio carries the personality and pacing of the game. Kiro: Development followed spec driven workflows so requirements, design, and task lists guided implementation and kept the full stack coherent under time pressure. What powers it: A React app for drawing and game flow, an Express API, vision for reading the canvas and context, and ElevenLabs for the voice layer, plus modes like hints, word of the day, and AI generated words for variety. Your canvas is the prompt. The AI answers out loud.

SoundPost turns your journal entry into a personalized AI-generated soundtrack — one card a day, so you never lose how a day felt. Most days dissolve by bedtime. SoundPost fixes that: type a few lines about your day, and ElevenLabs' music generation API produces a unique track that captures the mood. turbopuffer vector search on your journal embeddings then surfaces the days that "felt just like this" — so patterns in your life become visible over time. Tech stack: ElevenLabs — music generation API to synthesize the personalized daily soundtrack from extracted mood tags and key phrases turbopuffer — vector database for semantic similarity search across journal history

DramaGen is a full-stack web app where you sign in, type a calm, everyday message (for example, “please do your dishes”), choose a drama mode (rage, passive-aggressive, telenovela, disappointed parent, corporate fury, and others), and get back an over-the-top “rant” as playable, downloadable audio—not just text. Problem it solves: Most of us don’t actually want to send a real angry message, but we do want to vent or joke with friends in a way that lands. DramaGen turns “I’m too tired to argue” into a shareable audio clip that feels personal and funny instead of flat text. ElevenLabs: I use Instant Voice Cloning so users can record a short sample and hear the dramatic script in their own voice, plus Voice Design for preset character voices (chef, CEO, telenovela star, etc.). The final line is synthesized with ElevenLabs text-to-speech so every mode sounds like a performance. Replit: The project was built and deployed in the Replit ecosystem—fast iteration from idea to a published web app others can try, with secrets, hosting, and a workflow that matches how Replit is meant to ship voice-powered products end-to-end.

Eternal Jam Session is a deployed collaborative AI music room. You open a shareable URL, hold to talk or type directions (e.g. “more pads, slower, no vocals”), and the app turns that into the next section of music while updating mood, a command timeline, and stem-style visuals for everyone in the room. Problem it solves: Typical AI music tools feel like one-off generations with no shared history. Here the jam is persistent: the room stores prompt history in SQLite, can auto-evolve when it’s been quiet (scheduled idle pass), and broadcasts state so collaborators see the same mix version and timeline—instantly, without refreshing. Cloudflare: Workers serve the Vite/React UI and /api routes (new room IDs, multipart voice upload for STT, audio fetch). The core is a JamRoom Durable Object via the Agents SDK: @callable RPC from the browser, Workers AI (Llama) as a “director” that outputs structured JSON for the next composition, and schedules for idle evolution. ElevenLabs: Scribe v2 for speech-to-text, Music API (stream) for the stereo bed, with hooks for sound effects and stem separation when the director asks for layers or ear candy. API keys stay server-side on the worker.

AITAH?! turns Reddit AITA threads into an interactive “AI courtroom,” helping people quickly understand messy social conflicts by combining the original post, top comments, and a clear verdict. We used ElevenLabs Conversational AI (live Prosecutor, Defense, and Judge agents via WebSocket) for real-time debates and voice interaction, plus ElevenLabs simulateConversation + TTS with timestamps to generate narrated short-form reels; for this week’s partner, we used Firecrawl Search API as an agent tool (search_evidence) and to pull supporting “receipts” that enrich both the app and video output. What makes it special is the Reddit Jury vs AI Verdict dynamic: the system blends crowd opinion with agent reasoning, then packages the result into a social-ready 60-90s story reel, turning chaotic comment threads into something understandable, debatable, and shareable.