Enterprise Deployment System

From Idea to Production AI Infrastructure

11 production services, an agentic framework, a PostgreSQL-backed knowledge graph, and a multi-agent Memory Archive with 30+ MCP tools — designed, built, and operated solo from a Product Management background.

View Live Infrastructure →

Production Services

11 services running 24/7 — Career Bot, Memory Archive, Knowledge Graph, agentic framework, and more

From PM to Full-Stack AI Operator

Designed, built, and operate the entire system solo — from launchd service definitions and nginx upstreams to a multi-agent Memory Archive with 30+ MCP tools, a PostgreSQL-backed Knowledge Graph, and an agentic framework shared across services. Built from a Product Management background using AI-assisted development with disciplined architecture, CI, and operational rigor.

Product Management Multi-Agent Systems MCP / RAG / KG DevOps PostgreSQL

Career Bot — Public Chat

RAG-backed Q&A about my work and experience. Houses the production-live Talent Manager agent (protocol mirror + phase-gated state + rate-limited executor). Click to open the chat on the home page.

Flask · :8000 Claude RAG Talent Manager

Memory Archive + Knowledge Graph

Central brain for every agent. PostgreSQL-backed memories, instructions, sessions register, and a knowledge graph of nodes & edges. Exposes 30+ MCP tools at /memory-archive/mcp (bearer auth) for any Claude session to pull live context.

Flask · :5009 PostgreSQL MCP · 30+ tools Knowledge Graph

Resume Builder

Single-stage Opus pipeline that converts a job description + base résumé into a tailored, JSON-validated PDF. Hits Memory Archive for past-job context; uses Redis for job queue + caching.

Flask · :8004 Claude Opus Redis PDF gen

Blog (Public + Admin)

Two-service blog: a public reader and an authenticated admin that drives an AI content-generation pipeline. Persists posts and analytics in PostgreSQL; ships through the same deployment pipeline as everything else.

Flask · :8006/:8007 PostgreSQL AI authoring

Agentic Framework

Pip-installable package shared by every service: ProviderRegistry (multi-LLM), cost tracking, workflows, RAG, evaluation. One source of truth for how services talk to LLMs — change the registry, every service inherits the change.

Python package pip install -e Cost tracking RAG / Eval

Continuous Learning Agent

Weekly scanner that mines Memory Archive for stale memories, stale-but-important entries, and unused instructions. Auto-edits low-importance items; routes the rest to a proposals dashboard for human approval. Runs Sundays via launchd.

launchd PostgreSQL Slack pre-filter

Personal Assistant + Morning Brief

Flask + APScheduler daemon (port 5017) and a 6:30 AM digest service (port 5014) that stitches together Memory Archive activity, GitHub PRs, service health, and Slack approvals into one report.

APScheduler Slack launchd · cron

Prototype Studio + Discotheque

Internal sandbox for new ideas and a media-tinkering surface. Shares the same deploy pipeline, MA observability, and CI gates as the public services — proving that internal tools and production tools live on one substrate.

Flask · :5015 / :5016 Internal Same pipeline

Contact API + FAL App

Lightweight services that round out the public surface — a contact form endpoint to SMTP and an image-generation app via Fal.ai. Both wired into the same launchd + nginx + CSP infrastructure as the rest.

Flask · :8003 / :8001 SMTP Fal.ai

Deployment Pipeline

From local edit to live production — code path through the pipeline

System Status & Operations

Enterprise-grade infrastructure with professional monitoring

System Health

Fully Operational

Uptime

99.9% Availability

Security

SSL/TLS Enabled

Backups

Automated Daily

Technical Excellence & Innovation

Professional DevOps practices from first principles

Enterprise Operations

Built professional-grade deployment pipeline with automated safety mechanisms, health monitoring, and zero-downtime operations across multiple applications.

One-command deployment automation
Automated backup and recovery systems
Real-time health monitoring and alerts
Professional logging and analytics

VPS Infrastructure Mastery

Designed and deployed complete Virtual Private Server infrastructure from scratch, managing Ubuntu server configuration, service orchestration, SSL/TLS security, and multi-application hosting architecture.

Full VPS deployment and management
Multi-application service orchestration
SSL/TLS security with auto-renewal
Professional monitoring and alerting

Command Center Operations

Lightning-fast deployment and monitoring operations

< 30s

Full Backup & Deploy

1-Click

Git Commit & Push

Real-time

Health Dashboards

Instant

System Validation

Intelligent Monitoring

Real-time system dashboards providing instant visibility into application health, performance metrics, service status, and operational analytics across all production environments.

Live application health monitoring
Automated git workflow integration
Performance analytics tracking
Proactive system alerting

DevOps Pipeline Architecture

Professional deployment and monitoring infrastructure

Automated Deployment Pipeline

Safety Backup

Automated backup creation with timestamp to /var/backups/www-applications/

Code Synchronization

Rsync deployment from dev to production with exclusion filters

Permission Management

Automated chown operations ensuring proper www-data ownership

Service Restart

launchctl kickstart -k system/com.coreyscherrer.<svc> with a curl-loop until the health endpoint returns 200 — breaks throttle on stuck launchd jobs without dropping in-flight requests.

Health Verification

Automated health checks and deployment verification

Monitoring & Analytics

📊

Real-time Analytics

Per-service cost tracking via the agentic-framework registry. LLM spend, latency, and error rates feed the Memory Archive; Career Bot Q&A history persists to PostgreSQL.

🔍

Health Monitoring

Dedicated health endpoints with JSON status reporting and service validation

📝

Log Management

Centralized logging with journalctl integration and nginx access/error tracking

Security & Backup Strategy

🔒

SSL/TLS Security

Let's Encrypt certificates with auto-renewal and HTTPS enforcement

💾

Automated Backups

Pre-deployment safety backups with timestamp retention and centralized storage

🚨

Recovery Procedures

One-command rollback capabilities with documented recovery processes

Infrastructure Architecture

macOS launchd-managed services, nginx upstreams, PostgreSQL + Redis data tier

Development Environment (multi-agent worktrees)

🌳

Git worktrees per branch

/Users/admin/Sites stays on main; every active branch lives in its own /Users/admin/Sites-wt/<branch>. Prevents silent branch flips when multiple Claude agents work the repo concurrently.

📦

Per-service venvs + dev ports

Each service has its own venv/ and dev port (5000s range). Career Bot 5000, Resume Builder 5003, Blog 5006/5007, Memory Archive 5009, Prototype Studio 5015, Discotheque 5016 — all backed by the shared agentic-framework package.

🪝

Session hooks → Memory Archive

Every Claude Code session loads MA context on start (load-context.sh) and syncs memories on exit. Sessions register itself for auditability at /memory-archive/sessions.

🛡️

Per-service CLAUDE.md + AGENT_OF_RECORD.md

Each service ships its own instructions and an explicit owner agent. STOP-READ-FIRST headers fail-loud if context-loading hooks didn't fire — prevents agents from acting without knowing the rules.

Production Host (macOS + launchd + nginx)

🌐

Nginx (:80 → :443)

Reverse proxy with Let's Encrypt SSL termination, strict CSP (so vendor libs are self-hosted, no CDN scripts), and per-route auth gating for the Memory Archive admin surface.

⚙️

launchd (not systemd)

Every service runs as a com.coreyscherrer.<service> launchd job. Restart with sudo launchctl kickstart -k system/com.coreyscherrer.<svc>. Cron-style work (orphan audits, weekly learning agent, NAS backup) runs as scheduled launchd jobs.

🐘

PostgreSQL — single source of truth

One Postgres instance backs Career Bot, Blog, Resume Builder, Memory Archive, Knowledge Graph, and the Continuous Learning Agent. Database ownership documented in docs/architecture/DATABASE_OWNERSHIP.md so no service writes outside its lane.

⚡

Redis — queues + caching

Job queue for Resume Builder's PDF pipeline and rate-limit/state cache for Career Bot. Independent of Postgres so a slow LLM call never blocks transactional reads.

💾

3-2-1 backup tier

Local snapshots + Tailscale-mounted Synology NAS (DS918+) over SMB form the third leg of a 3-2-1 backup strategy. Daily NAS health-check posts to Slack.

Production Services (11 launchd jobs)

🤖

career-bot · :8000

Public chat. Houses Talent Manager (production-live multi-agent reference pattern).

🧠

memory-archive · :5009

PostgreSQL-backed memories + instructions + sessions + Knowledge Graph. MCP hub at /memory-archive/mcp with 30+ tools (bearer auth).

📄

resume-builder · :8004

Single-stage Opus → JSON → PDF pipeline. Uses Redis for job queue + caching.

📰

blog-public · :8006 blog-admin · :8007

Two-service blog: public reader + authenticated admin with AI authoring pipeline.

🎨

fal-app · :8001 contact-api · :8003

Image generation via Fal.ai and a contact-form-to-SMTP endpoint.

🛠️

prototype-studio · :5015 discotheque · :5016

Internal sandbox + media surface. Same pipeline, same observability as public services.

⏰

personal-assistant · :5017 morning-brief · :5014

APScheduler daemon and a 6:30 AM digest service that aggregates MA activity, GitHub PRs, service health, and Slack approvals.

Memory Archive + Knowledge Graph (the brain)

🔌

MCP endpoint — 30+ tools

/memory-archive/mcp exposes capture/search/update memory, instructions, knowledge-graph queries, sessions, and specialized review agents (architecture, database, security, observability, FinOps, UX, reliability) — bearer-auth protected, callable from any Claude session.

🕸️

Knowledge Graph — nodes & edges

Services, agents, decisions, learnings, and instructions are first-class nodes. Edges encode owns/depends-on/produces/relates-to. Visualized at /context-graph.html and /memory-archive/graph; agents query the KG to answer cross-service questions instead of grep-ing the codebase.

📋

Sessions register + audit

Every session writes session_id, agent_key, cwd, and hook_fired bool. Dashboard at /memory-archive/sessions shows fired/missed counts in a 48h window — cheap insurance that the context-loading discipline is working.

🔁

Continuous Learning Agent

Weekly scanner. Three deterministic detectors: stale memories, stale-but-important, unused instructions. Auto-edits low-importance items; routes the rest to the proposals dashboard with Slack approve/deny pre-filter.

Agentic Framework + Talent Manager pattern

🧩

agentic-framework — shared package

Pip-installable. ProviderRegistry abstracts Anthropic/OpenAI/Fal behind one interface. Built-in cost tracking, workflows, RAG, and evaluation. Imported by Career Bot, Resume Builder, Blog, FAL App — single source of truth for LLM access.

🪞

Talent Manager — protocol mirror + phase gates

Production-live agent inside Career Bot. Mirrors LinkedIn-style talent-search protocol; phase-gated state machine (scout → resolve → outreach) with rate-limited HTTP executor. Reference template that other service agents follow.

🛡️

Specialized review agents

Architecture, Database, Security, FinOps, Observability, Reliability, UX — each a standalone Opus 4.7 agent exposed as an MCP tool. Architecture Agent flags blast-radius and contract-at-risk; Database Agent gives APPROVE / APPROVE_WITH_CONDITIONS / BLOCK / NEED_INFO verdicts on schema diffs.

📑

Per-service AGENT_OF_RECORD.md

Every public service has an explicit owner agent and a CLAUDE.md drafted by the Documentation Agent (cited sources + explicit TODOs). When you ask "who runs X?" the answer is in the file, not in someone's head.

CI / Review / Deploy gates

✅

GitHub Actions — lint, build, tests

Every PR runs the same gates. Branch protection on main blocks direct push — everything lands via PR + squash merge.

🐰

CodeRabbit AI review

Automated review pass on every PR. Surfaces silent failures, type-design gaps, comment rot, and test-coverage holes before a human reads the diff.

🚀

dev-ops.sh deploy <service> --yes

Single command: rsync dev → prod, fix ownership, kickstart launchd, curl-loop until the health endpoint returns 200. shared/ deploys via a marker in requirements.txt that triggers deploy-common.sh automatically.

🌳

Cross-service health probe

After any deploy: probe ALL 11 services, not just the one touched. Catches regressions where a shared-package change breaks a service you didn't think you'd touched.

Service Architecture Diagrams

Explore detailed architecture for each microservice

Experience Professional AI Infrastructure

Explore the live production environment showcasing enterprise-grade deployment practices and AI application management.

Talk to my AI Career Bot View Live System Topology