I build and operate production AI systems.

Agent Hierarchy

Four autonomous agents operating with graduated authority. Each agent has scoped tool access — no agent can exceed its privilege boundary. The orchestrator routes intent, delegates tasks, and enforces the authority chain.

Production Stack

• Codex + Claude Code — autonomous code generation and PR review
• Agent Zero — self-improving autonomous workflows with FAISS vector memory
• Paperclip — multi-agent delegation with tool-level authorization
• Hermes — the implementation layer: intent routing, audit logging, gateway control

Runtime

Docker on Linux. Traefik routing. PostgreSQL for state. WebSocket gateways for real-time communication. Loading... continuous uptime with automated watchdog recovery.

Three-Layer Data Protection

• Layer 1 — Immutable Audit Trail: SHA-256 hash-chained. Every AI decision, every HITL approval, every data access — timestamped, hashed, and linked to the previous entry. Tampering with any entry breaks the chain. Verifiable at any time.
• Layer 2 — Consent Registry: Scope-based. Per-client. Per-purpose. Instant withdrawal with cascading data purge. Not a policy document — enforced in code at the gateway.
• Layer 3 — HITL Gates: Mandatory human approval for any write operation touching PII, any clinical decision, any financial transaction. Read ops are autonomous. Write ops require human sign-off.

Watchdog Monitoring

Automated cron-based watchdog monitors all services. Detects failures before clients notice. Exponential backoff with self-healing attempts. Escalates to human only when automated recovery fails — no alert fatigue.

Deployment Pipeline

• Source inventory — every dependency mapped before build
• Conflict log — inconsistencies flagged before code is written
• Missing context ledger — assumptions documented, not buried
• Duplicates report — existing work identified, nothing rebuilt unnecessarily

Every deployment inherits the compliance framework, the watchdog, and the audit trail. Build once. Benefit everywhere.

How Intent Classification Works

Deterministic classification layer sits between the user and the AI. Incoming messages are classified into one of 6 intent categories before any LLM inference. Each category has its own handler, its own authority scope, and its own tool access.

• Category A (Informational): Autonomous response. No tools needed.
• Category B (Booking): Read-only calendar access. Human confirms.
• Category C (Clinical): Read patient data. Zero write. HITL mandatory.
• Category D (Financial): Read billing status. Zero write. Human approval.
• Category E (Urgent): Bypass queue. Immediate human notification.
• Category F (Unknown): Flagged. Queued for human review.

The principle: the AI should never have more authority than the task requires. Graduated authority at the classification layer, not the model layer.