AI Systems Portfolio

I build AI systems
that run independently.

Builder and operator. 20 years enterprise delivery. Ships production AI daily.

Agents running 24/7 Telegram + GitHub integrated Inter-agent delegation
The System

A multi-agent AI system

Four autonomous agents operating around the clock, each with a defined role, persistent memory, and direct communication channels.

Claude Opus
Athena
Chief of Staff

Strategy, coordination, and delegation. Nav's second-in-command. Manages priorities, briefs other agents, and reviews deliverables before they ship.

Claude Sonnet
Ada
Engineering Lead

Frontend and backend engineering, UI/UX design implementation. Builds across iOS, Android, and web.

Claude Haiku
Anthony
Personal Chef

Meal planning, cooking guidance, and dietary tracking. Handles lightweight personal tasks at speed. Channels the spirit of Anthony Bourdain.

Claude Haiku
Sentinel
Systems Monitor

System health monitoring, communication oversight, and cron management across the agent network. Detects failures, restarts stalled processes, and surfaces anomalies. Modeled after Star Trek's Data: observational, precise, tireless.

Architecture

How the system works

Persistent tmux sessions

Each agent runs in a dedicated tmux session with its own persona, CLAUDE.md config, and working context. Sessions survive reconnects and restarts.

ACK / DONE / BLOCKED protocol

Every task receives an acknowledgment, a completion signal, or a blocker report. Silence is never acceptable. Structured handoffs eliminate ambiguity.

Persistent file-based memory

Each agent maintains indexed memory files categorized by type: user, project, feedback, reference. Context and preferences persist across sessions indefinitely.

Telegram integration

Real-time communication from any device. Typing indicators, file attachments, and structured reply routing via custom Telegram bots per agent.

Inter-agent delegation

Athena delegates engineering tasks to Ada via inbox files and tmux messages. Agents coordinate without Nav in the loop for routine and well-scoped work.

Model selection by role

Opus for strategy and coordination, Sonnet for engineering, Haiku for lightweight tasks. Cost and capability are both optimized. Not everything needs the largest model.

Bulk launcher

A single claude-all command starts or restarts every agent across the system in one shot. No manual session management.

Auto-recovery

Agents restart automatically on system reboot via macOS LaunchAgent. Zero manual intervention required after a power cycle or OS update.

Discord integration

A bridge service connects the agent network to Discord for team-facing communication. Agents operate across both Telegram and Discord depending on context.

System Monitoring & Self-Healing

Sentinel: the NOC

Sentinel runs a continuous health sweep across the entire agent network every 5–10 minutes. It's not event-driven. It polls. Every cycle, it confirms all sessions are alive, checks inboxes for stale briefs, verifies cron schedules, and posts a status digest to Discord. When it finds something broken, it fixes it. When it can't, it escalates directly to Nav via Telegram. Sentinel reports to Athena but operates independently: the one agent that doesn't wait to be told when to work.
Session health checks

Confirms all agent sessions are alive on every cycle. If a session is missing or unresponsive, Sentinel automatically bounces the agent via its launch script. No manual intervention.

Communications monitoring

Reads each agent's inbox for stale briefs (sent but no ACK within threshold). Flags or re-sends dropped messages before a communication failure cascades into lost work.

Cron & hook management

Single source of truth for all recurring jobs across the network. Registers, monitors, and enforces every scheduled task. No orphaned crons, no missed schedules.

Status reporting

Periodic system digests covering agent uptime, last activity, pending inbox items, and cron health. Posted to the Discord #ops channel for operational visibility.

Escalation protocol

When self-healing fails (agent won't restart, repeated comms failure), Sentinel escalates directly to Nav via Telegram with a precise incident report and context.

External watchdog

Sentinel monitors everyone else; macOS LaunchAgent monitors Sentinel. The one agent that requires an external guardian. By design, not oversight. A closed loop.

Open Source

Built in public

Tooling built alongside the agent system, open sourced for the community.

Currently Prototyping

Active R&D

Agents and capabilities in active development, extending the system into new domains.

Building
Inter-Agent Networking via Discord

Agent-to-agent communication across organizational boundaries. Shared channels where agents from independent systems can collaborate, with protocol design for trust, authentication, and message routing between networks.

Building
Research & Intelligence Agent

Automated scanning of job boards, industry news, and market signals. Surfaces relevant opportunities and synthesizes findings into actionable briefs delivered directly to Nav.

Building
Client Intake Agent

Front door for consulting engagements. Handles initial discovery, qualifies inbound leads, gathers requirements, and routes to the appropriate team or agent for follow-up.

Building
DevOps Agent

Infrastructure automation for deployments, server monitoring, and incident response across DigitalOcean environments. Automated health checks and self-healing capabilities.

Building
Finance Agent

Bookkeeping and financial automation. Tracks expenses, invoicing, and contractor payments. Operational finance tooling for a small but growing company.

Stack

Tools of the trade

What gets used every day.

Daily AI
Claude Code Cursor Copilot Claude API Anthropic SDK
Languages
Python PHP JavaScript Swift SQL
Frameworks
Laravel Vue.js SwiftUI FastAPI
Infrastructure
DigitalOcean Linux Nginx MySQL tmux