AI Game Master // Edge AI

ORACLE

The Ship Knows What You Did.

A fully offline AI game master running on a $15 Raspberry Pi Zero 2W. ORACLE drives SUNFALL: DRIFT -- a sci-fi social deduction tabletop game where players hunt void-corrupted cultists aboard a dying starship. Real-time dramatic narration, game state tracking, and atmospheric tension. No cloud. No internet. Just dread.

View All Projects

What Is It

A 1.3B parameter language model distilled from a 7B parent, quantized to Q2_K, and deployed on a Raspberry Pi Zero 2W. It acts as the AI narrator and game master for a physical tabletop game -- generating dramatic commentary, tracking game state, and responding to player actions in under a second.

What Is SUNFALL: DRIFT

A social deduction game set aboard the starship Meridian. The crew is fractured. Some are loyal. Some have been touched by the Void. Players vote, accuse, and sacrifice as the hull crumbles and ORACLE watches everything -- narrating the descent in real time.

Why It Matters

No game master needed. No app to download. No server to connect to. ORACLE runs entirely on a credit-card-sized board drawing 1.2 watts, producing 12-18 tokens per second with sub-second latency. The future of tabletop immersion fits in your pocket.

Hear The Ship Speak

Dramatic AI Narration, In Real Time

ORACLE doesn't just track numbers. It reads the game state -- hull integrity, crew suspicion, vote patterns -- and generates atmospheric narration that matches the tension at the table. Accusations feel heavier when the ship itself is watching.

  • Contextual narration based on live game state
  • Tracks hull integrity, crew status, vote history
  • Responds to RFID card scans in under 800ms
  • Escalating tension as the hull degrades
  • Unique voice lines for accusations, votes, and betrayals
oracle-v1 // SUNFALL: DRIFT
// round 4 -- hull at 34%
SCAN: RFID_07 -- player "Kessler" casts ACCUSE on "Voss"
ORACLE:
"The accusation hangs in the recycled air.
Voss has voted correctly three times.
But the hull remembers what the crew forgets.
Sector 7 was sealed after Voss ran diagnostics alone."
Latency: 680ms
Tokens: 47 @ 14.2 tok/s
Hull: 34% -- critical threshold
// LED ring pulses red. BLE speaker delivers line.

SUNFALL: DRIFT

A social deduction game set at the edge of oblivion

"The Meridian was 14 months from port when the Void found us. It didn't break the hull. It broke the crew. Now every vote is a prayer, every accusation a gamble, and the ship's AI watches it all with something that looks almost like hunger."

The Void Corrupted

  • Hidden cultists sabotage the ship from within
  • Loyal crew must identify and eject them
  • The Void spreads each round if cultists survive
  • Trust is the most dangerous resource

Vote or Die

  • Each round ends with an accusation and vote
  • Wrong votes damage the hull further
  • ORACLE tracks every decision and remembers
  • The ship's narration shifts based on who you trust

The Meridian

  • Hull integrity is the game clock
  • Systems fail as damage accumulates
  • ORACLE's narration grows more desperate
  • At 0% hull, the Void wins -- and everyone loses

No Two Games Alike

  • AI narration is generated live, never scripted
  • ORACLE references past votes and betrayals
  • Different role assignments each game
  • Emergent drama from player interaction + AI

How ORACLE Thinks

Knowledge distillation, extreme quantization, and edge deployment -- from 7B parameters to a $15 board.

Stage
What Happens
Teacher Model
7B parameter model generates high-quality game narration examples
Knowledge Distillation
Student model (1.3B params) learns to replicate the teacher's dramatic voice
Q2_K Quantization
1.3B model compressed to 1.2GB with 2-bit quantization via llama.cpp
Context Injection
Game state (hull %, votes, roles) injected as structured prompt prefix
Edge Inference
llama.cpp runs natively on ARM Cortex-A53 at 12-18 tok/s, 380MB RAM

Benchmarks

Real performance numbers from a $15 single-board computer

12-18 Tokens/sec sustained inference
<800 ms First Token time to first output
380 MB RAM of 512MB available
1.2 Watts total power draw
1.2 GB Model Q2_K quantized
$15 Hardware Cost Pi Zero 2W
1.3B Parameters distilled from 7B
0 Cloud Calls fully offline

The Hardware Stack

Everything ORACLE needs to run a game fits on a tabletop

Compute & Display

Pi Zero 2W

1GHz quad-core ARM, 512MB RAM

CrowPanel Display

Game state & narration output

LED Ring

Atmospheric feedback & alerts

1.2W Total

USB-C or battery powered

Input & Output

RFID Reader

Player card scanning for actions

BLE Audio

Wireless speaker for narration

Python Engine

Game logic & state management

llama.cpp

Optimized ARM inference

🛰

Fully Offline & Portable

No WiFi. No app store. No subscription. Just shuffle the deck and power on.

Instant On

Boot to game-ready in seconds. No updates, no login, no pairing. Plug in power and ORACLE awakens.

🌐

Zero Internet

Every token generated locally on the Pi. Play in a cabin, a basement, a convention hall -- anywhere.

$15 Brain

The entire AI runs on a Raspberry Pi Zero 2W. No GPU, no server rack, no cloud bill. Fifteen dollars.

Under The Hood

The inference pipeline, from card scan to spoken narration.

oracle-pipeline
# 1. RFID scan triggers event
IN rfid_reader.scan() -> player_action
# 2. Python engine updates game state
ENG game.process(action, hull=34, round=4)
# 3. State injected into prompt template
CTX prompt.build(state, vote_history, roles)
# 4. llama.cpp generates narration
LLM llama_inference(prompt) -> 47 tokens @ 680ms
# 5. Output to display, speaker, LEDs
OUT display.render() | ble.speak() | led.pulse(RED)

The Stack

Raspberry Pi Zero 2W llama.cpp Q2_K Quantization Knowledge Distillation Python RFID BLE Audio CrowPanel LED Feedback ARM Cortex-A53 Edge AI Social Deduction

The Ship Is Listening

A 1.3-billion-parameter narrator running on hardware that costs less than a pizza. No cloud. No game master. No mercy. ORACLE turns every accusation into theatre and every vote into a crisis. The Meridian is waiting for its next crew.

Explore All Projects →