build-small-hackathon/phantom-grid

🕵️ Phantom Grid

An AI-driven noir detective game. You are a detective hunting a phantom suspect across a stylized London grid. Issue notices, raise lookouts, set blockades, and interview AI-roleplayed witnesses whose memories decay over time — all rendered in a custom gr.Server HTML/JS board interface

Track: Delightful — Thousand Token Wood (an AI-driven game).

🏅 Prize-category badges

• —🎮 Thousand Token Wood / Delightful — a playable, AI-driven detective game.
• —🎨 Off Brand — fully custom gr.Server HTML/JS frontend, well beyond the stock components.
• —🪶 Small & Mighty — runs entirely on a single under-32B model (MiniCPM4.1-8B).

🤖 Model & inference

• —Model: `openbmb/MiniCPM4.1-8B` (text, bf16 transformers) — ~8B params, well under the 32B cap.
• —Inference: in-process Hugging Face transformers, placed on cuda at module load (using ZeroGPU's PyTorch CUDA emulation), with the real GPU attached only inside a @spaces.GPU-decorated generate() call.

🖥️ Hardware

Runs on ZeroGPU (NVIDIA RTX Pro 6000 Blackwell, large / 48 GB VRAM; 40 min/day for Team org members). Each generation is capped at PHANTOM_GRID_ZEROGPU_DURATION seconds (default 90). No voice path for now (PHANTOM_GRID_WITNESS_CHAT_TTS=0).

🎥 Demo video

https://www.youtube.com/watch?v=p8iSjatInXo

📣 Social post

Launch post on X: https://x.com/unityashtv/status/2066633879109382378

▶️ How to play

Start a new case, read the briefing, then use the board tools (notices, lookouts, blockades, searches) and question witnesses to corner the suspect before the turn limit.

Local Deployment

For running a version of the game with a locally running llama cpp backend find the code at https://github.com/U4AR/JohnDoe