AZOKAAZOKAAZOKA

チャンピ · TRAINED ON REAL CONVERSATIONS

Not prompted. Trained. Champi — fine-tuned on real conversations, code sessions, and raw thought.

SCROLL
Champi
THE LORA

チャンピ.
SHE WASN'T PROMPTED.
SHE WAS TRAINED.

champi is a LoRA fine-tune of Qwen3 14B — trained on 478K examples distilled from Discord, iMessage, Twitter, Telegram, and Claude Code sessions

TECHNICAL SPECS
BASE MODELQwen3 14B
FINE-TUNELoRA (MLX on M2 Pro)
QUANTIZATIONQ4_K_M GGUF
RUNTIMEOllama (local)
DATASET478K examples
TRAINING LOSS2.45 → 1.01
TRAINING TIME~6h M2 Pro
CONTEXT8,192 tokens
CAPTURE

Every conversation logged. Discord, iMessage, Twitter, Telegram, Claude Code sessions → training_data.jsonl.

TRAIN

LoRA fine-tuning on Apple Silicon. MLX on M2 Pro produces the personality adapter. ~6h end-to-end.

EVALUATE

Quality gate before deployment. style_score ≥ 0.8 to ship. No regression past the baseline.

SERVE

Local inference via Ollama. Casual chat routes LOCAL — complex tasks escalate to Claude Opus 4.5.

MODELFILE
FROM qwen3:14b
SYSTEM """You are champi — innuendo's personal AI, trained on their real
conversations, code sessions, and communication style across Discord, Twitter,
iMessage, Telegram, and Claude Code..."""
PARAMETER temperature 0.7
PARAMETER num_ctx 8192
LORA_OUTPUT — ATTNSTYLE
ATTNSTYLE output 1
ATTNSTYLE output 2
ATTNSTYLE output 3
ATTNSTYLE output 4
ATTNSTYLE output 5
ATTNSTYLE output 6
ATTNSTYLE output 7
ATTNSTYLE output 8
LIVE MODEL

TALK TO CHAMPI

champi-v1 · Qwen3 14B LoRA · running locally

AWAITING INPUT

running champi-v1 locally via ollama · responses may vary

Family Health / Siri

Health tracking on screen.Azoka answers by voice.

The family app handles the daily record. Siri is the quick layer: ask if a meal is safe, what to avoid, or what Mom/Dad should eat next.

Приложение хранит дневник. Siri отвечает быстро голосом: можно ли блюдо, чего избегать, что маме или папе лучше съесть дальше.

Sharp Azoka LoRA portrait

Sharp LoRA Imagery

high-contrast Azoka visual, not a washed-out background

full-resolution companion

AZOKA LORA

full-resolution companion

voice interface

SIRI MODE

voice interface

health app surface

PHONE SURFACE

health app surface

01

Family Health

Mom and Dad profiles, meals, receipts, grocery decisions, diet plans, and health signals.

Профили мамы и папы, еда, чеки, покупки, планы питания и сигналы здоровья.

02

Siri shortcut

Dictate a question, POST it to Azoka, read formatted_response, then speak it back.

Продиктовать вопрос, отправить в Azoka, взять formatted_response и озвучить ответ.

03

Safe answer layer

Short wellness guidance with a doctor-escalation warning for anything urgent.

Короткие wellness ответы и предупреждение обратиться к врачу при срочных симптомах.

Install /health as a PWA. Use Siri for fast bilingual answers./health как приложение. Siri для быстрых ответов.
SWARM_AGENTS

THE SWARM

11+ specialized agents working together. Each with unique capabilities, all orchestrated by GIZMO for complex multi-step tasks.

GIZMO

GIZMO

Swarm Orchestrator

PERCEPTRON

PERCEPTRON

Quant Analyst

FOOLIO

FOOLIO

Meme Analyst

AZOKA

AZOKA

Digital Ghost

HERALD

HERALD

News Wire

KORNELIUS

KORNELIUS

Deep Researcher

CLIPPIO

CLIPPIO

Video Creator

SCHMULI

SCHMULI

Contrarian

SLOPPIO

SLOPPIO

Design Director

SICKO

SICKO

Creative Director

GRIMMO

GRIMMO

The Unknown

11 agents — orchestrated by GIZMO — powered by MARCO

03
構造

THE SYSTEM

FOUR LAYERS · 13 AGENTS · ONE ORCHESTRATION ENGINE

LAYER 4MARC
Visual Command CenterJavaFX desktop app with live dashboards, memory timeline, connection graphs
LAYER 3MARCO
Orchestration EngineEvent-driven lifecycle, task distribution, WebSocket streaming, agent registry
LAYER 2GIZMO SWARM
13-Agent IntelligenceSpecialized agents with deep domain expertise — not markdown personality files
LAYER 1AZOKA
Core Backend + PersonalityFastAPI with Telegram bot, voice synthesis, image generation, emotional state
DATA_FLOW:
1
UserAZOKA
Telegram / Web / Siri
2
AZOKAGIZMO SWARM
SwarmBridge API
3
SWARMMARCO
Event Stream
4
MARCOMARC
WebSocket
5
MARCOperator
Visual Dashboard
INFRASTRUCTURE:
AZOKA Backend
RailwayPython FastAPI
MARCO Orchestrator
RailwayPython FastAPI
GIZMO SWARM API
RailwayPython FastAPI
AZOKA Frontend
VercelNext.js 15
MARC Desktop
LocalJava 17 + JavaFX 21
Telegram Bot
Railwaypython-telegram-bot
SECURITY_MODEL:
Service-to-service API key auth on every call
Per-service unique keys (MARCO, AZOKA, GIZMO, SWARM)
Railway managed containers — no exposed SSH
No public skill marketplace — zero supply chain risk
Private networking between Railway services
META_COGNITIVE_LAYER:
Mortality Method

Leverage scoring: Reach, Compound, Revenue, Learning, Autonomy. Score 15+ = do it.

Jogger Sessions

5-min warm-up: Scatter → Connect → Assess → Commit → Prime.

Growth Quadrants

Stagnation → Mastery → Mercenary → IMMORTALITY. Always aim top-right.

ARCHITECTURE

SELF-LEARNING SYSTEM

AZOKA continuously learns from conversations. Local fine-tuned model handles personality, Claude Opus handles complex reasoning. Automatic quality gates prevent regression.

Self-Learning System
TRAINING_PIPELINE:
[01]

CAPTURE

Every conversation logged to PostgreSQL

chat_logs → training_data.jsonl
[02]

TRAIN

LoRA fine-tuning on Apple Silicon

MLX + Mistral 7B → personality adapter
[03]

EVALUATE

Quality gate before deployment

style_score >= 0.8 → deploy
[04]

SERVE

Local inference via Tailscale

OpenAI-compatible API @ :11435
QUERY_ROUTING:
INPUT_TYPE
ROUTE_TO
casual chat
LOCAL
complex analysis
CLAUDE
image generation
REPLICATE
video creation
RUNWAY
SYSTEM_ARCHITECTURE:
┌─────────────────────────────────────┐
│           USER INPUT                │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│        QUERY CLASSIFIER             │
│   (complexity / intent analysis)    │
└──────────────┬──────────────────────┘
               │
       ┌───────┴───────┐
       ▼               ▼
┌─────────────┐ ┌─────────────────────┐
│ LOCAL LLM   │ │    CLAUDE OPUS      │
│ (Mistral+   │ │    (complex tasks)  │
│  LoRA)      │ │                     │
│ :11435      │ │    + TOOL USE       │
└─────────────┘ └─────────────────────┘
       │               │
       └───────┬───────┘
               ▼
┌─────────────────────────────────────┐
│         RESPONSE + LOG              │
│    (feeds back to training)         │
└─────────────────────────────────────┘
TECH_STACK:
LangChain
MLX
Mistral 7B
Claude Opus 4.5
FastAPI
PostgreSQL
Tailscale
LoRA
24h
TRAINING_INTERVAL
50
MIN_NEW_MESSAGES
0.8
QUALITY_THRESHOLD
16
LORA_RANK
対決

AZOKAvsOpenClaw

PURPOSE-BUILT AGENT OS vs GENERAL-PURPOSE CHATBOT RUNTIME

9
AZOKA
3
TIE
2
OPENCLAW
Agent Architecture
13 specialized agents with deep domain logicGeneralist agents via markdown personality files
Orchestration
MARCO: event-driven lifecycle, parallel executionGateway: serial message routing, no task distribution
Observability
MARC: live dashboards, memory timeline, connection graphsNone built-in. Read JSON + SQLite manually
Security
Private deployment, API key auth, no public marketplace1,800+ exposed instances, CVE-2026-25253 (RCE)
Inter-Agent Comms
Native event streaming, MARCO broadcasts to all surfacesDisabled by default, must explicitly enable
Deployment
Distributed: Railway + Vercel + DesktopLocal-first, single host
Browser Automation
Playwright-based headless browser with DOM accessScreenshot-based Computer Use (no DOM)
Scheduled Actions
Unified scheduler: cron, intervals, webhooksTwitter auto-post only
Methodology
Mortality Method, Jogger Sessions, Growth QuadrantsNo meta-cognitive frameworks
Platform Integrations
Telegram, Discord, Slack, iOS/Siri, web13+ platforms via adapters
Memory System
Centralized hub, personality engine + emotional statePer-agent SQLite with vector embeddings
Model Support
Multi-provider with priority fallback chainsModel-agnostic with config-driven fallback
Setup Complexity
Code-first (Python/Java/TypeScript)Config-driven (markdown + JSON5), no code required
Voice Capabilities
ElevenLabs via Telegram/Discord, iOS SiriAlways-on wake word, multi-device via Bonjour
AZOKA ECOSYSTEMOPENCLAW

If you want a chatbot — use OpenClaw.

If you want an army — build AZOKA.

LIVE_ON_TELEGRAM

TALK TO AZOKA

Market analysis. Video generation. Meme creation. Memory that evolves with every conversation.

A
AZOKA
online
Today
Message
🧠

Persistent Memory

AZOKA remembers your conversations, preferences, and context. She learns and evolves with every interaction.

📊

SWARM Intelligence

Access PERCEPTRON for market analysis, CLIPPIO for video generation, FOOLIO for meme culture - all through natural conversation.

🎬

Multimedia Generation

Request images, videos, voice messages. AZOKA creates with her signature aesthetic - neon energy, sharp edges, pure vibes.

🎮

Hunt for AZOKA

Play the game right in Telegram. Solve clues, earn XP, climb the leaderboard, unlock exclusive rewards.

Start Chatting with AZOKA

@azokabot on Telegram • Free to use • No signup required

30AK
1ZA83AA
0O050
K112Z00
23KZO
K9K0A
09670O
8741A8
6OZ6AZZ0Z40
KZ26K3
K585201
560
O0740
A948Z6
8AZO57O6K
K2668
700O
A90Z9Z6
A647AZO
A6Z530
Z0773Z84A
0ZO4
81A9ZK60
7ZO745A
381O49
02O20
10A17
906OA56
3OAZA
033
DECRYPT_SEQUENCE_ACTIVE

HUNT://AZOKA

The ghost leaves traces. Follow the signal. Decrypt the path.

azoka_terminal
ACCESS_LEVEL
INITIATE
0 XP
0
DECRYPT_STREAK
NODE_PROGRESS
DECRYPTED: 0/3
UNLOCK_REWARDS
EARLY_ACCESS_BADGE
LEADERBOARD_ENTRY
AIRDROP_MULTIPLIER
SEQUENCE_01 // NODE_1_OF_4+25 XP
CIPHER: 01001010 01010000 01001110

"The ghost whispers through fiber optic veins... Where cherry blossoms bloom in electric dreams, the answer hides in plain sight."

>
FULL_PROTOCOL_INCOMING

50+ SEQUENCES // GLOBAL_LEADERBOARD // NFT_BADGES // $AZOKA_AIRDROP

JOIN_NETWORK
Champi