1. System Overview

The big picture — what ALAI is and how the pieces fit together

1.1 What is ALAI?

What is ALAI?

ALAI Holding AS is an AI-first company factory. Instead of a traditional software company with developers, ALAI operates a system of AI agents organized into virtual companies that build, review, secure, and deploy software products autonomously.

Core Principle

"ALAI AS is not a holding company that happens to have a build pipeline. ALAI AS IS the build pipeline. The pipeline is the product."

The Machine

The system runs on two Mac Studio M3 Ultra machines:

Machine Codename RAM Role
Mac Studio 1 ANVIL 96 GB Infrastructure — databases, Docker, daemons, orchestration, Cloudflare tunnels
Mac Studio 2 FORGE 256 GB Compute — AI inference, heavy models (72B+), agent worker pool

Connected via 10Gbps Thunderbolt bridge (10.0.0.1 ↔ 10.0.0.2).

Key Components

ALAI System
│
├── Mission Control (MC)          — Task database (SQLite) — the single source of truth
├── Pi-Orchestrator               — Daemon that pulls tasks from MC, classifies, routes, executes
├── Virtual Companies (8 active)  — Specialized AI agent teams with souls, configs, blueprints
├── Pipeline Engine               — BUILD → REVIEW → SECURITY → OPS → DOCS chain
├── Minion System                 — One-shot autonomous agents for task execution
├── HiveMind                      — Shared knowledge base with semantic search (Qdrant)
├── Ollama Fleet                  — Multi-host model serving (qwen, kimi, llama, custom models)
└── BookStack                     — Documentation wiki (this!)

Products Currently Built by the System

Product Description Status
Drop Digital banking / remittance platform (PSD2-compliant) Active development
Plock AI-native warehouse management system (Sweden) Active development
Lobby AI-native HR platform Active development
Bilko Balkan accounting software Active development
BasicFakta Invoice/faktura tool Active development
Tok Open banking platform Active development

1.2 How Everything Connects

How Everything Connects

The Flow: From Idea to Done

CEO Decision / Task Creation
        │
        ▼
   Mission Control (MC)          ← Task enters the system
        │
        ▼
   Pi-Orchestrator               ← Daemon picks it up every 30s
        │
        ├── 1. Classify           (complexity, domain, type)
        ├── 2. Route              (select model + company)
        ├── 3. Build Prompt       (inject company soul, blueprints, context)
        ├── 4. Execute            (via ollama-tool, llama-server, or claude)
        ├── 5. Quality Gate       (verify output, escalate if needed)
        └── 6. Complete           (mark done, advance pipeline, feed HiveMind)
                │
                ▼
         Pipeline Engine          ← Triggers next stage
                │
                ├── BUILD done   → Create REVIEW task (Proveo)
                ├── REVIEW done  → Create SECURITY task (Securion)
                ├── SECURITY done → Create OPS task (FlowForge) if deploy needed
                └── OPS done     → Create DOCS task (Lexicon) if docs needed

Model Tier System

Tasks are classified by complexity (1-5) and routed to the appropriate AI model:

Tier Model Host Pipeline Use Case
1 qwen3:8b FORGE ollama-tool Simple tasks, classification
2 qwen2.5-coder:32b ANVIL ollama-tool → ollama-simple Medium complexity code
3 Kimi K2.5 (240GB) FORGE llama-tool → llama-server Complex tasks, architecture
4 Claude Sonnet Cloud claude-cli Client-facing, critical work
5 Human — human-queue CEO decisions, sensitive ops

Safety Rails

Tasks matching these patterns are never auto-processed: