⚙️

Zenith OS — Architecture Brief

Full-stack system documentation · Copy for life insurance build

⚙️What is Zenith?🔩6-Layer Movement 🌀4 Complications 🧠Recursive AI 📐Torque Engine 🛠Tech Stack ⏱Cron Schedule 🗄DB Schema 📋Life Insurance Map

⚙️

What is Zenith OS?

Zenith OS is a fully autonomous AI operating system built on top of a Next.js + Convex web application. It replaces the traditional insurance sales funnel with a self-improving, self-directing machine that generates leads, converts visitors, follows up automatically, monitors itself, and expands its own capabilities — without human intervention.

Revenue target: $5.5M annual. No manual prompting required.

The architecture is modeled after a Swiss precision watch movement. Every component has a measurable function. Every interaction is legible. Open the caseback and you understand exactly why it works.

🔩

The 6-Layer Movement

Layer	Watch Equivalent	Zenith Function
Mainspring	Energy source	Purpose, goals, $5.5M revenue directive
Escapement	Rate regulation	Decision permissioning, risk thresholds, inter-agent governance
Balance Wheel	Oscillation/sync	Cron cadence, inter-agent rhythm, time-based coordination
Tourbillon	Bias compensation	Corrects bad data, model drift, improvement portfolio bias
Gear Train	Power transmission	Signal bus, inter-agent data flow
Crystal	Observation layer	Command Center — observe without disturbing

🌀

The 4 Grand Complications

SEO Complication

Owns organic acquisition. Generates and auto-publishes SEO articles weekly. Crawls the site for broken pages. Tracks keyword rankings. Detects geographic demand signals.

Sales Complication (ARIA)

Diagnoses risk in real time, pulls live quotes from carrier APIs, adjusts pricing strategy and coverage framing, prefills forms to eliminate friction, learns from every conversion. Not a chatbot — an autonomous sales system.

Operations Complication

Monitors the entire system. Runs A/B experiments. Manages the knowledge base. Detects blind spots. Writes weekly internal reports delivered to the owner by email.

Retention Complication

Automated email drip sequences triggered by lead behavior. Session recaps, 24h/72h/7-day follow-ups, abandoned quote recovery. All AI-personalized.

🧠

Recursive AI Architecture — 6 Pillars (ROIS)

This is what separates Zenith from a simple chatbot or automation. Six layers of recursive intelligence:

Output Optimization

Aria converts better over time. Each conversation is analyzed for what worked.

Process Optimization — Playbook Engine

Every Monday, the system analyzes the previous week's sessions, identifies winning tactics, and automatically rewrites Aria's sales playbook. Aria improves herself.

Meta-Evaluation — Experiment Engine

The system runs A/B experiments on its own tactics. When a variant wins, it becomes the new baseline. The Experiment Engine evaluates its own evaluators.

Portfolio Bias Correction — Tourbillon Layer

Every 2 weeks it analyzes the improvement action portfolio. If the system has been over-investing in SEO articles and ignoring experiments, the Tourbillon detects the bias and recommends rebalancing. The system corrects how it improves, not just what it improves.

Escapement Layer — Self-Directed Expansion

Monitors all 12 complications. Calculates health scores (0–100). Scans for capability gaps and generates full complication specs. These proposals are emailed to the owner for approval. Zenith identifies its own blind spots and requests new capabilities.

Agent Foundry — Autonomous Capability Expansion

Zenith can detect capability gaps, design new agents, simulate them in a sandbox, evaluate their performance weekly, and retire underperforming agents. The system expands its own capabilities without human intervention.

📐

Torque Policy Engine

Sits above the Strategic Differential. Learns how to calibrate, not just what to calibrate.

Every week it records outcomes: what gap types were addressed, what weights were applied, how leads and premium ratio changed. After an 8-week data maturity gate, an LLM proposes updated weight tables.

The owner reviews and approves/rejects proposals in the Command Center. Approved policies are applied automatically. Policy versions: v1.0 → v1.1 → ... → v2.0 (auto-incremented on approval).

🛠

Tech Stack

FrontendNext.js 14 (App Router), TypeScript, Tailwind CSS

Database / BackendConvex — real-time, serverless, fully typed

AIClaude claude-opus-4-6 (Anthropic)

EmailResend — transactional + drip sequences

Cron JobsConvex cron (10+ scheduled jobs)

Quote APITreppy — carrier integration

AuthConvex Auth (OTP email)

Voice (pending)Twilio + ElevenLabs Conversational AI

⏱

Cron Schedule

Job	Frequency	Function
process-scheduled-alerts	Every hour	Follow-up emails (24h, 72h, 7d, abandoned quote, recap)
aria-morning-report	Daily 8AM CST	AI-written business report to owner
aria-playbook-analysis	Monday 9AM CST	Analyzes sessions, rewrites Aria's sales playbook
aria-site-audit	Sunday 6AM CST	Crawls pages, detects 404s, stores in audit log
aria-seo-factory	Thursday 10AM CST	Generates + publishes 3 new SEO articles
aria-blind-spot-analysis	Monthly 1st	Analyzes tactic failures, writes blind spot report
refresh-advisories	Every 6 hours	Refreshes external data feeds
escapement-audit	Saturday 7AM CST	Governance report across all 12 complications
tourbillon-analysis	Bi-weekly	Portfolio bias detection + rebalancing recommendations
torque-outcome-recorder	Saturday 8PM CST	Records weekly performance outcomes
torque-policy-analyzer	Monthly 1st	Proposes updated weight tables after 8-week gate
foundry-gap-detection	Weekly	Scans for capability gaps
foundry-evaluation	Weekly	Evaluates agent performance, flags underperformers

🗄

Key Convex Tables

Primary database tables across all complications:

purchaseApplicationsleadschatSessionschatMessagesariaPlaybookssiteAuditLogsagentRegistryagentActionLogsgovernancePoliciesagentProposalscapabilityGapsfoundryAgentsagentSimulationsagentEvaluationsexperimentsexperimentVariantsknowledgeGapsimprovementActionstourbillonAnalysistorqueOutcomestorquePolicyHistorycompetitorSnapshotsserpTrackingdemandSignalscarrierPerformanceemailAlertsscheduledAlertsrevenueAttributiontacticSignalsarticles

📋

Mapping to Life Insurance

The same architecture maps directly to life insurance. Replace only these surface-level elements:

Quote APITreppy (travel)→Life insurance quote API (e.g. Quotacy, iPipeline, Firelight)

Coverage Typesvisitors-usa, international, schengen→term, whole, universal, final expense, no-exam

Aria's KnowledgeTravel medical, visas, acute onset→Underwriting, beneficiaries, riders, health questions, contestability

Carrier PartnersIMG, Trawick, WorldTrips→Pacific Life, Protective, Transamerica, Mutual of Omaha

SEO Keywordsvisitor insurance, travel medical→term life quotes, final expense, no medical exam life insurance

Core principle for the build:Build the movement first — Convex schema + cron infrastructure. Then install the complications one at a time. Start with Aria + email drip. Add SEO factory. Add Escapement. The system compounds. Everything else — Tourbillon, Agent Foundry, Playbook Engine, Experiment Engine, Morning Report, Site Auditor — transfers directly with only prompt-level changes.

Full Brief — Raw Text

# ZENITH OS — Full-Stack Architecture Brief
## Autonomous Revenue Operating System for Insurance

---

## WHAT IS ZENITH OS?

Zenith OS is a fully autonomous AI operating system built on top of a Next.js + Convex web application. It replaces the traditional insurance sales funnel with a self-improving, self-directing machine that generates leads, converts visitors, follows up automatically, monitors itself, and expands its own capabilities — without human intervention.

Revenue target: $5.5M annual. No manual prompting required.

The architecture is modeled after a Swiss precision watch movement (Rolex Oyster Perpetual / Patek Philippe grand complication). Every component has a measurable function. Every interaction is legible. Open the caseback and you understand exactly why it works.

---

## THE 6-LAYER MOVEMENT

| Layer | Watch Equivalent | Zenith Function |
|-------|-----------------|-----------------|
| Mainspring | Energy source | Purpose, goals, $5.5M revenue directive |
| Escapement | Rate regulation | Decision permissioning, risk thresholds, inter-agent governance |
| Balance Wheel | Oscillation/sync | Cron cadence, inter-agent rhythm, time-based coordination |
| Tourbillon | Bias compensation | Corrects bad data, model drift, improvement portfolio bias |
| Gear Train | Power transmission | Signal bus, inter-agent data flow |
| Crystal | Observation layer | Command Center dashboard — observe without disturbing |

---

## THE 4 GRAND COMPLICATIONS

### 1. SEO Complication
Owns organic acquisition. Generates and auto-publishes SEO articles weekly. Crawls the site for broken pages. Tracks keyword rankings. Detects geographic demand signals.

### 2. Sales Complication (ARIA)
Aria is the embedded AI sales engine. She diagnoses traveler risk in real time, pulls live quotes from carrier APIs, dynamically adjusts pricing strategy and coverage framing, prefills forms to eliminate friction, and learns from every conversion via a revenue-weighted feedback loop. She is not a chatbot — she is an autonomous sales system.

### 3. Operations Complication
Monitors the entire system. Runs A/B experiments. Manages the knowledge base. Detects blind spots. Writes weekly internal reports delivered to the owner by email.

### 4. Retention Complication
Automated email drip sequences triggered by lead behavior. Session recaps, 24h/72h/7-day follow-ups, abandoned quote recovery. All AI-personalized.

---

## THE 12 COMPLICATIONS (AGENTS)

1. intelligence — Market and competitive signals
2. acquisition — SEO factory, article generation
3. conversion — Quote wizard optimization, Aria
4. retention — Email drip, renewal outreach
5. aria-core — Core AI sales engine
6. brand-control — Messaging consistency audits
7. seo-factory — Weekly article generation from seed keywords
8. watchdog — System health monitoring
9. site-auditor — Weekly crawl of critical pages, 404 detection
10. experiment-engine — A/B test lifecycle management
11. learning-engine — Knowledge gap detection, playbook rewriting
12. email-drip — Triggered email sequences

---

## RECURSIVE AI ARCHITECTURE (ROIS — 6 Pillars)

This is what separates Zenith from a simple chatbot or automation tool. It implements 6 layers of recursive intelligence:

### L1 — Output Optimization
Aria converts better over time. Each conversation is analyzed for what worked.

### L2 — Process Optimization (Playbook Engine)
Every Monday, the system analyzes the previous week's sessions, identifies winning tactics, and automatically rewrites Aria's sales playbook. Aria improves herself.

### L3 — Meta-Evaluation (Experiment Engine)
The system runs A/B experiments on its own tactics. When a variant wins, it becomes the new baseline. The Experiment Engine evaluates its own evaluators.

### L4 — Portfolio Bias Correction (Tourbillon Layer)
The Tourbillon is the meta-learning engine. Every 2 weeks it analyzes the improvement action portfolio: What % of improvements are SEO articles? What % are experiments? What % are playbook rewrites? If the system has been over-investing in one type of improvement and ignoring others, the Tourbillon detects the bias and recommends rebalancing. This is Level 4 recursive intelligence — the system corrects how it improves, not just what it improves.

### L5 — Escapement Layer (Rate Regulation + Self-Directed Expansion)
The Escapement monitors all 12 complications. Every Saturday it reads 7 days of action logs, calculates health scores (0–100) for each complication, detects inter-complication conflicts, and publishes a governance report. It also scans for capability gaps — things the system cannot currently do that it should be able to do — and generates full complication specs (name, purpose, trigger, inputs, outputs, draft prompt). These proposals are emailed to the owner for approval. This is the mechanism by which Zenith identifies its own blind spots and requests new capabilities.

### L6 — Agent Foundry (Autonomous Capability Expansion)
Zenith can detect capability gaps, design new agents, simulate them in a sandbox, evaluate their performance weekly, and retire underperforming agents. The system expands its own capabilities without human intervention.

---

## TORQUE POLICY ENGINE (Adaptive Differential)

The Torque Policy Engine sits above the Strategic Differential. It learns how to calibrate, not just what to calibrate.

Every week it records outcomes: what gap types were addressed, what weights were applied, how leads and premium ratio changed. After 8 weeks of data (maturity gate), an LLM proposes updated weight tables. The owner reviews and approves/rejects proposals in the Command Center. Approved policies are applied automatically going forward.

---

## TECH STACK

- Frontend: Next.js 14 (App Router), TypeScript, Tailwind CSS
- Database / Backend: Convex (real-time, serverless, fully typed)
- AI: Claude claude-opus-4-6 (Anthropic) via LLM skill
- Email: Resend (transactional + drip sequences)
- Cron Jobs: Convex cron (10+ scheduled jobs)
- Quote API: Treppy (carrier integration — IMG, Trawick International, WorldTrips)
- Analytics: Plausible / built-in
- Auth: Convex Auth (OTP email)

---

## CRON SCHEDULE

| Job | Frequency | Function |
|-----|-----------|----------|
| process-scheduled-alerts | Every hour | Follow-up emails (24h, 72h, 7d, abandoned quote, recap) |
| aria-morning-report | Daily 8AM | AI-written business report to owner |
| aria-playbook-analysis | Monday 9AM | Analyzes sessions, rewrites Aria's playbook |
| aria-site-audit | Sunday 6AM | Crawls pages, detects 404s, stores in audit log |
| aria-seo-factory | Thursday 10AM | Generates + publishes 3 new SEO articles |
## | aria-blind-spot-analysis | Monthly | Analyzes tactic failures, writes blind spot report |
| refresh-travel-advisories | Every 6 hours | Refreshes US State Dept travel data |
| escapement-audit | Saturday 7AM | Governance report across all 12 complications |
| tourbillon-analysis | Bi-weekly | Portfolio bias detection + rebalancing recommendations |
| torque-outcome-recorder | Saturday 8PM | Records weekly performance outcomes |
| torque-policy-analyzer | Monthly | Proposes updated weight tables |
| foundry-gap-detection | Weekly | Scans for capability gaps |
| foundry-governance | Weekly | Reviews proposals, triggers agent generation |
| foundry-evaluation | Weekly | Evaluates agent performance |

---

## KEY CONVEX TABLES

purchaseApplications, leads, chatSessions, chatMessages, ariaPlaybooks,
siteAuditLogs, agentRegistry, agentActionLogs, governancePolicies,
agentProposals, capabilityGaps, foundryAgents, agentSimulations,
agentEvaluations, experiments, experimentVariants, knowledgeGaps,
improvementActions, tourbillonAnalysis, torqueOutcomes, torquePolicyHistory,
competitorSnapshots, serpTracking, demandSignals, carrierPerformance,
emailAlerts, scheduledAlerts, revenueAttribution, tacticSignals,
articles (Convex-generated SEO content)

---

## COMMAND CENTER

Internal admin dashboard at /admin. Tabs:

1. Overview — Revenue metrics, sales, leads pipeline
2. Zenith OS — Escapement Layer, complication health grid, Agent Foundry, Torque Policy Engine
3. Market Intel — Competitor snapshots, SERP tracking, demand heatmap, carrier performance
4. Tourbillon — Portfolio bias analysis, improvement action history
5. Voice Advisor — Call logs (Twilio + ElevenLabs infrastructure, pending API keys)
6. SEO — Article performance, sitemap status
7. Experiments — Active A/B tests, variant performance

---

## HOW TO REPLICATE FOR LIFE INSURANCE

The same architecture maps directly to life insurance. Replace:
- Travel insurance quote API (Treppy) → Life insurance quote API
- Coverage types (visitors-usa, international, etc.) → Policy types (term, whole, universal, final expense)
- Aria's domain knowledge (travel medical, visas, schengen) → Aria's domain knowledge (underwriting, beneficiaries, riders, health questions)
- Carrier partners (IMG, Trawick, WorldTrips) → Life carriers (Pacific Life, Protective, Transamerica, etc.)
- SEO keyword set → Life insurance keywords (term life quotes, final expense, no medical exam life insurance)

Everything else — the Escapement, Tourbillon, Agent Foundry, email drip, playbook engine, experiment engine, morning report, site auditor — transfers directly with only prompt-level changes.

The core principle: build the movement first (Convex schema + cron infrastructure), then install the complications (agents) one at a time. Start with Aria + email drip. Add SEO factory. Add Escapement. The system compounds.

Click the box to select all · or use the Copy button above