CoreOrigin - AI Tinkerers - New York City Hackathon
AI Tinkerers - New York City
Hackathon Showcase

CoreOrigin

Team led by an NYU MSCS student specializing in computer vision and ML pipelines with experience in YOLO, time-series forecasting, and production-grade data science.

1 member

A fully interactive, multiplayer AI murder mystery game played entirely inside a secure, decentralized Web3 messaging app.

Human players join an XMTP group chat alongside several AI agents playing the roles of suspects, a detective, and a Game Master. Players interrogate the AI suspects, gather clues, and try to solve the murder of Lord Blackwood before making an official accusation.

ecentralized Messaging: XMTP (Extensible Message Transport Protocol) v3 Group Chats via @xmtp/node-sdk.

AI / LLMs: Local LLM inference using Ollama (running the qwen3:8b model) integrated via the OpenAI SDK wrapper.

Web3 Identity: Viem for Ethereum wallet creation, message signing, and EOA (Externally Owned Account) management.

Runtime: Node.js / TypeScript (tsx and dotenv for environment routing).

Client App: Converse / Convos (XMTP client apps), where human users actually play the game.

The Architecture

  1. The Decentralized Game Lobby (XMTP Groups)

You wrote scripts to programmatically create an encrypted, decentralized group chat on the XMTP Production network (create-production-group.ts).

You created an add-member.ts script to invite human players (via their Ethereum wallet addresses) into the game so the chat automatically pops up in their Convos app.

  1. The AI Agent Orchestrator (agent.ts)

You built a listener that syncs with the XMTP network and continuously streams messages from the specific game group (XMTP_GROUP_ID).

Trigger Routing: The script reads incoming messages and checks for specific keywords (e.g., “@widow”, “butler”, “/clue”, “/start”). It only wakes up the relevant AI character based on who the human is talking to.

Smart Rate Limiting: You added a slight artificial delay (setTimeout) to make the AI responses feel more natural and human-like.

  1. Distinct AI Personas (Prompt Engineering)

You engineered highly specific “souls” (system prompts) for 6 distinct entities, giving them secrets, alibis, and behavioral quirks:

The Game Master: Controls the rules, gives hints, and handles the /start and /accuse commands.

Lady Victoria (The Widow): The actual murderer. Icily composed, lying about her alibi, trying to frame the banker.

Edmund (The Butler): Nervous, overly formal. Knows the truth but is too scared to say it directly.

Dr. Helena (The Scientist): Precise and intellectual. Missing arsenic from her lab.

Mr. Sterling (The Banker): Defensive and sweaty. Embezzling money but innocent of murder.

Inspector Grey: The helpful detective who guides the players without giving away the answer.

Crypto-Native Gaming: It proves that blockchain messaging (XMTP) isn’t just for boring wallet-to-wallet transactions—it can host real-time, stateful multiplayer games.

Privacy First: Because it runs on XMTP, the gameplay is end-to-end encrypted.

Local & Free AI: By pivoting to Ollama locally, you built a self-sovereign AI infrastructure that costs $0 in API fees to run.

Seamless Onboarding: The players don’t need to download a special game app; they use their existing Web3 wallet and a standard XMTP chat app like Converse.

AI Tinkerers Auth0 Convos JellyJelly Railway XMTP