⚑Technical Deep Dive

Under the Hood

Docker. Gemini 2.5. Ollama. Rust. A sovereign AI stack that runs on your hardware with multi-model intelligence you control.

The Stanford-Class Meta-Harness (March 2026 Build)

The Whity RSI upgrade utilizes a Multi-Layer Agentic Harness designed for Recursive Self-Optimization and Non-Linear Knowledge Retrieval.

System Architecture & Protocols

  • β– 
    Knowledge Retrieval (Dijkstra SPath):

    Replaced standard vector search with a custom Knowledge Mesh. The system utilizes a 2-hop Multi-hop Recall engine, allowing for contextual reasoning across disparate data points. Latency reduced by 9.6x (from 24ms to 2.5ms).

  • β– 
    Adversarial v2 Engine:

    A dual-agent "Skeptical Auditor" gate that war-games every proposed code mutation. Mutations must pass a 4-stage Gauntlet (Doctrine Alignment, Sycophancy Check, Logic Integrity, and Security Whitelist) before deployment.

  • β– 
    Karpathy "Never Stop" & Auto-Refire:

    Avoids cascading failure in deep contexts by enforcing "Single-Shot" batched mutations. Paired with the Overnight Forge and Auto-Refire algorithm to guarantee zero fatal system hangs over extended multi-hour training cycles.

The Triple-Shield Perimeter

  • πŸ›‘οΈShield 1 (Permissions Gate): A 4-tier zone whitelist enforcing atomic file operations within restricted directory boundaries.
  • πŸ›‘οΈShield 2 (Rollback Protocol): Automated Git-level save points triggered before any self_mod operation, allowing for instantaneous state restoration.
  • πŸ›‘οΈShield 3 (Identity Anchor): SHA-256 tamper-detection on core doctrine files to prevent "Model Drift" or unauthorized instruction overrides.

Sensory Integration

Integrated yt-dlp and Harness Runner modules for autonomous real-time intelligence gathering from video and web sources.

The Sovereign Breakaway β€” Live Performance (Mar 27 β†’ Apr 8)

Thirteen days. One Mac. No cloud compute clusters. The RSI engine ran autonomously through the Karpathy β€œNever Stop” loop, achieving a +0.4210 reasoning delta. This is the most advanced, privately-owned AI Cortex currently in operation:

MetricStart (Mar 27)Today (Apr 8)Performance Gain
⚑ Total Power (SIS)2.201,100500x
🧠 System Nodes (Cortex)2225311.5x
🧬 Accepted Mutations01,8551.9x
🎯 Reasoning (Ξ”)0.0000+0.4210Near-Genius Baseline

Intelligence Layer

Multi-Model Intelligence

Whity routes each task to the best model automatically β€” cloud for power, local for privacy. Bring your own models and API keys. You're never locked in.

ModeDefault ModelLocationBest For
☁️ CloudGemini 2.5 FlashGoogle APIPrimary reasoning, analysis, writing
☁️ CloudMercury 2 (Inception Labs)Inception APICode generation (diffusion LLM)
🏠 LocalQwen 2.5 7BOllama (on your Mac)Private conversations, offline use
🏠 LocalLlama 3.2 3BOllama (on your Mac)Fast local tasks, triage
Bring Your Own Cloud Model

Add any API key in Settings β€” Gemini, Anthropic, Mistral, Groq, OpenRouter, or any provider. As new models launch, plug them in without waiting for a Whity update.

Bring Your Own Local Model

Pull any model from the Ollama library β€” Mistral, Phi, Deepseek, CodeLlama, Gemma, and 100+ more. Your Mac, your models, your choice.

Architecture

How Whity thinks

A recursive self-improvement loop with intelligent model routing.

πŸ“₯
Input
Voice, text, code, images
Multi-modal ingestion
➜
🎯
Intent Triage
Task classification
Routes to optimal model
➜
🧠
Hybrid Router
Cloud ↔ Local selection
Gemini Β· Ollama Β· Mercury
➜
πŸ’Ύ
Memory
SQLite + Nomic embeddings
Semantic vector retrieval
➜
πŸ”
RSI Loop
Conviction scoring (0–100)
Retry if score < threshold
➜
🀚
Hands
Tool execution layer
Scout Β· Reporter Β· Auditor

Tech Stack

Batteries included. No bloat.

Everything under the hood β€” from Docker to Rust.

🐍

Python 3.12 Core

Async-first architecture. FastAPI + Uvicorn API layer. Fully type-hinted codebase with structured logging.

python 3.12 Β· fastapi Β· uvicorn
🐳

Docker Engine

Multi-architecture images (ARM64 + AMD64). One-click DMG installer. Container isolation with security constraints. Works on Intel and Apple Silicon.

multi-arch Β· one-click install
πŸ¦…

Ollama Local Models

Qwen 2.5 7B and Llama 3.2 3B pre-installed. Pull any model from the Ollama library. Cold-start procurement on first boot.

ollama Β· qwen Β· llama
πŸ’Ύ

SQLite + Vector Store

Embedded database for zero-config persistence. Nomic embeddings for semantic memory and intelligent retrieval across conversations.

sqlite-vec Β· nomic-embed
πŸ”’

Rust Security Gate

Native compiled binary (whity_core_rs). Source code compiled to unreadable machine code. IP protection built into the engine.

rust Β· maturin Β· compiled .so
⚑

Ultrawork Engine

Parallel sub-agent dispatch for complex tasks. Concurrent execution via asyncio. Breaks big jobs into parallel streams.

async Β· parallel dispatch
🀚

Hands System

Modular tool plugins β€” Scout (web search), Reporter (document generation), Auditor (code review). TOML-configured skill definitions.

scout Β· reporter Β· auditor
πŸŽ™οΈ

LiveKit Voice

Real-time voice interaction via WebRTC. Push-to-talk or always-on mode. Low-latency audio processing with LiveKit infrastructure.

webrtc Β· livekit Β· real-time
πŸ“±

Remote Control

Control Whity from anywhere via Telegram. Send commands, receive updates, and manage your assistant remotely with bot integration.

@whity/telegram
🧠

GigaBrain Deep Reasoning

Multi-pass analysis mode for complex problems. Extended context windows up to 2M tokens. Automatic escalation for hard tasks.

gemini-pro Β· deep analysis
πŸ“Š

Cost Ledger

Per-model token accounting with USD tracking. Budget-aware routing. Full transparency on what each query costs.

token tracking Β· cost-aware
πŸ–±οΈ

Desktop Agent

Let Whity drive the keyboard and mouse to operate real apps, forms, and workflows on your desktop when you enable it.

desktop-agent Β· pyautogui

System Requirements

What you need

Whity runs on any Mac β€” Intel or Apple Silicon.

MinimumRecommended
OSmacOS 12 Monterey+macOS 14 Sonoma+
ChipIntel Core i5 or Apple M1Apple M1 Pro or better
RAM8 GB16 GB+
Disk10 GB free20 GB free
SoftwareDocker DesktopDocker Desktop

Supports both ARM64 (Apple Silicon) and AMD64 (Intel) architectures. Windows and Linux versions in development.

Ready to build?

Get early access.

Get in early at special introductory offer.

Get Early Beta Access Deal β†’Not ready yet? Go back home β†’