𓂀 Architecture

Multi-model orchestration powering The Sphinx

┌─────────────────────────────────┐ ┌──────────────────────────────────┐ │ Browser (Chrome) │ │ FastAPI Server │ │ │ │ │ │ index.html + app.js + style.css │ │ main.py │ │ │ │ ├── api/routes.py │ │ ┌─ Web Speech API (STT) ──┐ │ │ ├── game/engine.py (rules) │ │ │ Voice → text → question │ │ POST │ ├── game/judge.py (fuzzy match) │ │ └──────────────────────────┘ ├──────►│ ├── game/characters.py (data) │ │ │ │ ├── llm/interpreter.py │ │ ┌─ Audio playback ─────────┐ │ JSON │ │ ├─ RuleMatcher (regex) │ │ │ ElevenLabs TTS stream │◄──┤◄──────┤ │ ├─ IntentClassifier │ │ └──────────────────────────┘ │ │ │ └─ LLM QuestionInterpreter │ │ │ │ ├── llm/voice.py (Sphinx mood) │ │ ┌─ Dual interface ─────────┐ │ │ ├── llm/full_llm_handler.py │ │ │ Chat mode / Voice mode │ │ │ ├── tts/client.py (ElevenLabs) │ │ └──────────────────────────┘ │ │ └── store/memory.py (sessions) │ └─────────────────────────────────┘ └──────────────────────────────────┘ │ ┌──────────────────────────────────┼──────────────────┐ │ │ │ ┌─────▼──────┐ ┌──────▼───────┐ ┌──────▼──────┐ │ Mistral AI │ │ ElevenLabs │ │ Amazon │ │ │ │ │ │ Bedrock │ │ Large: Sphinx│ │ Flash v2.5 │ │ │ │ personality │ │ TTS voice │ │ Claude 3.5 │ │ + answers │ │ streaming │ │ Llama 3.1 │ │ │ │ │ │ Cohere R+ │ │ Small: Intent│ │ "Adam" deep │ │ Nova Pro │ │ classifier │ │ mysterious │ │ │ └──────────────┘ └──────────────┘ │ AI Opponents│ └─────────────┘ ┌──────────────┐ │ MCP Server │ │ │ │ 7 tools for │ │ AI agents to │ │ play the game│ │ via protocol │ └──────────────┘

Tech Stack

𓁹 The Sphinx Brain

Mistral Large interprets free-form questions, generates theatrical responses with mood-aware personality, and guards the secret character's identity.

mistral-large-latest

𓂀 Intent Classifier

Mistral Small runs fast intent classification — is it a question, a guess, or something else? Sub-200ms pre-filter before the main LLM.

mistral-small-latest

𓆣 Voice of the Sphinx

ElevenLabs streams a deep, mysterious TTS voice in real-time. The Sphinx speaks with gravitas.

eleven_flash_v2_5

⚡ AI Opponents

Four AI personas powered by Amazon Bedrock compete against human players in duel mode, each with a unique strategy and personality.

Bedrock Converse API

𓊗 MCP Server

AI agents can play LeSphinx programmatically through 7 MCP tools. Connect via stdio or SSE transport.

Model Context Protocol

𓃭 Truth Engine

Deterministic answer resolution from structured character data. The LLM never decides what's true — only how to say it.

RuleMatcher + FactStore

Game Flow

1
Player asks a question by voice or text
Web Speech API → FastAPI
2
RuleMatcher tries regex patterns for common attributes
Deterministic, <1ms
3
Intent Classifier determines: question vs guess
Mistral Small, ~150ms
4
Question Interpreter parses the question into structured attributes
Mistral Large, ~400ms
5
Answer Resolver checks against character facts deterministically
FactStore lookup
6
Sphinx Voice generates a theatrical, mood-aware response
Templates + Mistral Large fallback
7
ElevenLabs streams the Sphinx's voice to the player
TTS streaming, ~300ms first byte

API Endpoints

Method	Path	Description
`POST`	`/game/new`	Create a new game session
`POST`	`/game/{id}/ask`	Ask a yes/no question
`POST`	`/game/{id}/guess`	Guess the character
`POST`	`/game/{id}/hint`	Request a hint
`GET`	`/game/{id}/state`	Get current game state
`POST`	`/game/{id}/ai_turn`	Execute AI opponent's turn
`GET`	`/ai/personas`	List AI opponent personas
`GET`	`/leaderboard`	View top scores & stats
`POST`	`/leaderboard`	Submit score

Character Database

161 curated famous characters across 14 themes, 3 difficulty levels. Each character has structured attributes, facts, aliases, bilingual summaries, and photos.