Master

Horror is hard to design. Anyone can make a player jump. Few can make them dread. Most horror games chase scope — sprawling maps, hours of mechanics, scares on a timer. The dread fades. The player adapts. Tension dies in playtime. The Hollow House goes small on purpose. It is a fifteen-minute first-person horror experience. Three friends take a shortcut through the woods and stumble into a house that does not want them to leave. They solve riddles posed by the house itself, examine documents, listen to whispered hints from each other, and survive whatever is hunting them — before they become the next set of scratched-out faces on the family photographs. Atmosphere comes first. The house breathes through audio. Every character has a voice. Aarav, the nervous calm one. Priya, who hears it before anyone else. Sam, who is too quiet. The narrator is dry. The presence is none of those things. Lines are generated dynamically from a single dialogue script, then routed by character. The entity passes through a Presence bus — layered pitch-shifted playback through lowpass, distortion, and reverb — so it feels diegetically wrong, not just loud. Footsteps duck under speech. Heartbeat and whispers ramp as you descend. A CRT shader sits over every frame — scanlines, vignette, faint flicker — like found footage you should not be watching. The entity only spawns in the basement. That one rule turns every other room into anticipation. Under the hood: Godot 4 on the gl_compatibility renderer for portability. Every voice line is generated and cached via the ElevenLabs Text-to-Speech API across five distinct voices — including a custom-tuned narrator and a layered demonic mix for the presence. The closing credit is pre-generated in parallel with the ending so the audio lands instantly. Built entirely in Zed. Every GDScript, every scene tweak, every gameplay system was iterated through Zed's AI agent — spec, draft, refine, ship — at the pace of conversation. The editor was the workflow, not just the surface. The Hollow House does not chase scares. It earns them.

Submitted 30 Apr 2026

VoiceBoard — Live AI Meeting Co-Pilot with Persistent Memory VoiceBoard is a full-stack AI meeting co-pilot that listens to your meetings in real-time, transcribes everyone (including other participants via WebRTC audio capture), answers questions mid-meeting with spoken voice responses, and builds persistent memory across every session you've ever had. The Problem: After back-to-back meetings, critical decisions blur together. "Did we approve the budget?" "What did the client say about the timeline?" Teams waste hours scrolling chats, re-reading notes, or worse — guessing. What We Built: - Real-time transcription with speaker identification - Mid-meeting voice Q&A — ask "What did we decide about pricing?" and hear the answer spoken back via ElevenLabs TTS - Sentiment analysis heatmap showing the emotional arc of every conversation - AI-extracted action items, key decisions, and follow-ups - Multi-language live translation (9 languages) - Voice cloning — meeting summaries spoken in your own voice - Meeting minutes export (Markdown) - Cross-meeting memory — ask questions across months of meeting history - Chrome extension that works inside Google Meet, MS Teams, or any browser tab - Standalone web app at voiceboard.ksdas1245.workers.dev Cloudflare Stack (deeply integrated): - Durable Objects — persistent state per meeting room, real-time multi-user sync via WebSocket - Vectorize — every transcript line embedded and stored for semantic search across all meetings - Workers AI — Whisper for audio transcription, BGE for embeddings, Llama for Q&A/extraction - Workers — edge compute for the entire backend, globally distributed ElevenLabs Integration: - Text-to-Speech (eleven_flash_v2_5) — voice answers to mid-meeting questions, 60-second audio briefings - Voice Cloning — clone any participant's voice, deliver summaries in their voice Chrome Extension: included in repo (/extension folder)