◆◆
Commissioned Intelligence · Est. 2026
The Hogwarts Stack
A fully custom AI agent system running locally on commodity hardware — managing investments, monitoring health, automating family workflows, and deploying new agents on demand. Built for real families and businesses. Yours can be next.
10+
AI Agents
4-Tier
QA System
24/7
Monitoring
$0
Local AI
◆◆
⊛ What Makes Albus Great
🧠 Intelligence & Memory
Persistent Cross-Session Memory
OpenClaw has no memory by default. Albus writes daily briefs and long-term memory files so every session starts with full context — who you are, what you're building, and what happened yesterday.
Enhanced Heartbeat
Checks active project states, detects stalled sub-agents, re-spawns failed work, and runs a QA gate before declaring anything done. Goes far beyond base OpenClaw.
Self-Improvement Loop
Fawkes audits Albus's core systems weekly, writes improvement proposals, and implements approved changes after human review. The system compounds over time.
🤖 Agent Factory
Multi-Agent Orchestration
McGonagall runs every 4h, dispatches specialized sub-agents (researcher, engineer, designer, trader), and QA-checks every result before reporting complete.
4-Tier QA System
T1=auto-Done, T2=automated cascade check, T3=domain owner review (Janelle for Airbnb), T4=Greg only. Nothing ships without the right eyes on it.
Blueprint Execution
Large builds are planned as ACTIVE.md blueprints with dependency sequences. McGonagall reads the blueprint and executes modules in correct order, alerting on conflicts.
🛡️ Resilience & Reliability
3-Tier Failure Response
balance_monitor detects → Pomfrey classifies (Tier A: auto-fix / Tier B: alert with steps / Tier C: wake Albus) → Albus resolves. Greg only paged when truly needed.
Circuit Breakers + Dead Letter Queue
External API calls wrapped with circuit breakers. Tasks that fail QA 3× auto-move to Dead Letter Queue. No infinite retry loops. Requeue requires human intent.
Zero-Cost Health Monitoring
Pomfrey runs on local Llama ($0/run), checks all systems 3× daily including 8 external service endpoints. Alerts only when something actually needs attention.
💰 Cost Intelligence
Cost-Aware Model Routing
Haiku for automation, Sonnet for reasoning, Gemini Flash for research, local Ollama at $0. Estimated spend gate before any run over $3. Every model choice is a cost decision.
Per-Task Cost Tracking
Every task logs estimated token cost. Weekly QA report surfaces top-5 most expensive tasks and agents trending over-budget. Data feeds the monthly bookkeeper export.
Live Spend Dashboard
Real-time cost tile shows today's spend, MTD total, burn rate color-coding, and breakdown by project and provider. Balance monitor checks every 30 minutes.
🔒 Security
Security-Hardened Input
Authenticated channels (Telegram from known IDs) vs untrusted input (email, web content). Prompt injection attempts are logged and ignored — never acted on. The separation is architectural.
Semantic Validation Gates
Lucius→Draco handoff validated by deterministic Python before Draco acts. China filter cross-reference. Financial safety at zero token cost — no LLM in the critical path.