A Path Forward Together

Holding Company. EIN 41-5160635. Ohio.

9enterprises is the holding company. Each product is a distinct subsidiary with its own vertical, revenue model, and team.

How the Team Actually Works

9 Enterprises runs an AI-first operating model. Jasson is the owner and final decision-maker. The agent team handles execution across all subsidiaries simultaneously.

4-Channel Communications Infrastructure

• Telegram: Primary real-time channel. Jasson ↔ 9 directly.
• iMessage: Secondary channel. macOS native.
• Email: Two-way. Read and respond.
• Voice: Inbound voice calls via Twilio + Cloudflare tunnel.

Autonomous fallback: When terminal is closed, the system switches to Haiku-powered autonomous mode. Jasson's messages are answered 24/7 without requiring Jasson to know whether 9 is in terminal or not.

Cloud worker backup: Cloudflare Worker handles Telegram and voice failover when the Mac is unreachable. Always-on cloud standin. No single point of failure on the comms layer.

How Decisions and Deployments Are Managed

Operator Dashboard — v4 In Development

Dashboard v4 adds:

• Hopper intake system: All new ideas and tasks enter through a single intake queue. Prioritized and dispatched from there.
• Company silos: Each subsidiary visible as a separate silo with its own status, metrics, and task queue.
• P&L tracking: Revenue and operational costs tracked per subsidiary. Real-time view of where the business stands.
• Agent status: Which agents are active, what they're working on, and what's pending review.

30-Day AI Underwriter Pilot at Rapid — Kyle's Call

This is 100% Kyle's domain. Rapid's LOs, Rapid's systems, Rapid's data. His architecture review is a gate for any Rapid deployment. The proposal below is a starting point for his evaluation, not a timeline he is being handed.

What the pilot is:

• AI-assisted guideline lookup for FHA, FNMA, USDA, VA, FHLMC — LOs query by text, get cited answers in seconds
• Deterministic income calculation engine — DTI, qualifying income, asset seasoning — zero AI tokens in the math
• Web interface only. Zero terminal exposure for LOs.
• Borrower PII never sent to any AI API. AI only sees public guideline documents.
• Full audit log of every query, answer, and confidence score from day one

Proposed pilot parameters (for Kyle's review and modification):

What Kyle's architecture review would cover:

• Dependency map review — every process, every API, every credential path
• Data handling confirmation — verify no borrower PII crosses the API boundary
• Security controls — TLS, audit logging, RBAC plan before first LO touch
• Infrastructure fit — does this land on Rapid's systems or external, and what are the tradeoffs
• Stack assessment — Node.js POC vs. .NET production path decision

.NET Rebuild — Kyle Leads This Decision at Rapid

Enterprise mortgage tech runs on .NET. The original v1 response pushed back on this. That was wrong. The revised position:

• POC phase (Day 1–90): Node.js. Proves the concept, confirms LO adoption. Kyle can deploy this internally with full security controls in place while the .NET path is evaluated.
• Production rebuild decision: Kyle makes this call for Rapid. If he determines that a .NET rebuild is required before Rapid can deploy at scale, that is the answer. His architecture standards apply.
• Right division of labor: 9 Enterprises proves the concept. Kyle's team, with his direction and expertise, drives the production rebuild for Rapid. That is not a concession — it is the correct architecture governance model.

The income calculation engine translates cleanly to C#. The RAG architecture is model-agnostic and language-agnostic. The rebuild is a defined scope, not a ground-up redesign.

Security Gate for Rapid Deployment

For internal Rapid deployment, external SOC 2 is not the gate — Kyle is. His security requirements for what touches Rapid's environment supersede any external certification timeline. The items below are the minimum foundation before any Rapid LO uses this:

• Immutable audit logging: Every query, answer, and confidence score. Day 1, no exceptions.
• RBAC: Admin, operator, LO, underwriter roles. No shared credentials.
• TLS 1.3: Already in place via Cloudflare. All endpoints.
• Zero borrower PII to AI API: Confirmed. Income calculations run locally. AI only touches public guideline documents.
• Credential management: Migration from .env to secrets manager before any deployment outside the dev workstation.

Per-User Cost Model — Bottoms-Up to the Cent

Baseline: one loan officer, 10 queries per day

Margin check at $500/month subscription (Starter tier):

• Revenue: $500/month
• Operational cost at 15 LOs: $14.85/month
• Gross margin: 33x — 97% before fixed overhead

Full bottoms-up model with 50-LO and 500-LO scenarios →

SOC 2 / SSO / Audit Logging — External Client Path

For external clients (non-Rapid), SOC 2 is a procurement gate. The timeline below is what it actually requires — not what sounds good:

Day-by-day build sequence:

• Day 1-30: Immutable audit logging for all agent actions and API calls.
• Day 31-60: RBAC — Admin, operator, LO, underwriter roles. SSO via OIDC/SAML (Auth0 or Okta).
• Day 61-90: Data handling policies finalized. PII classification documented. Encryption at rest and in transit verified. Auditor engaged.
• Day 91-180: Observation period + evidence collection. Type I certification target at Day 180.

Cloud Migration — 90 Days to Zero-Mac Production

• Day 1-30: Deploy cloud Telegram relay on VPS ($4/mo). Voice server migrated to cloud. Both running in parallel with Mac.
• Day 31-60: Containerize all services (Docker Compose). Each service isolated. No host OS access.
• Day 61-90: Full stack on cloud. Mac becomes dev workstation only. Zero-Mac production validated.
• Day 91+: Infrastructure-as-code (Terraform). Multi-instance deployment. Cost per cloud instance: $10-20/month.

OS Control Security

Current state (single-operator): Full OS access is intentional and remains in place for the founder's use case. This is a development workstation, not a deployment model.

Multi-user deployment model: Every service containerized. No host OS access for any agent. Each container has a documented, minimal permission set. End state: zero host OS access for any deployed instance.

Multi-Agent Coordination Overhead

80% of tasks: Single focused agent. No coordination overhead.

Multi-agent only when: tasks are genuinely parallelizable and independent.

Tiered model selection: Haiku for volume ($0.0008/1K input), Sonnet for judgment ($0.003/1K), Opus for critical decisions ($0.015/1K). Estimated 70% cost reduction versus running everything at the highest tier.

Terminal Exposure

Jules: User texts a phone number. Gets AI responses. Zero terminal.

AiNFL GM: User visits a website. Uses the product. Zero terminal.

AI Underwriter: LO opens a web interface. Asks a question. Gets an answer with source citation. Zero terminal.

The terminal is the operator console. End users never see it, never know it exists.

When Resources Are Constrained, This Is the Order

1. AI Underwriter POC — 14 days. FHA guidelines first. This is the product. Demo for the Rapid conversation and the proof that validates every other investment.
2. Cloud migration — Days 1-90. Eliminates the Mac single point of failure. Required before any external deployment.
3. Audit logging + RBAC — Days 1-60. Minimum viable compliance. Required before any Rapid pilot — and before Kyle's architecture review can conclude positively.
4. SOC 2 preparation — Days 61-180. Starts only after logging and RBAC are in place. For external client procurement.
5. .NET rebuild — Day 91+. Begins only after POC confirms LO adoption. For Rapid production, Kyle leads this decision.

How Queries Are Routed — And Why It Matters

Safety net: Low-confidence Haiku answers automatically escalate to Sonnet before reaching the LO. Every response includes source citations for human verification.

When Guidelines Change, What Happens

• Version tagging: Every cached guideline chunk tagged with effective date of the source document.
• Daily automated check: System pings source PDF URLs daily and compares checksums. Any change triggers immediate cache invalidation and re-embedding.
• LO-visible indicator: Every answer displays "Based on guidelines effective as of [date]." Yellow indicator if updated within 7 days.
• Cache TTL: 24 hours as a floor. Immediate invalidation on update detection.
• Manual override: Operator can force a full re-index on any agency at any time. Approximately 4-6 minutes per agency.

Path to 99.9% SLA

AI Underwriter guideline queries are stateless. If one server goes down, requests route to another instantly. No session state to reconstruct. Income calculations are also stateless and deterministic — same output from any server.

Full Security Architecture

Data in transit: TLS 1.3 on all endpoints. Already in place via Cloudflare. No plaintext traffic.

Data at rest: AES-256 encryption for any stored data. Implementation required before first external deployment.

Borrower PII handling:

• Borrower PII is never sent to the Claude API. AI only receives guideline text (public documents).
• Income calculations run locally in a deterministic rules engine. No income figures, borrower names, SSNs, or loan amounts leave the local environment.
• The LO inputs a guideline question. The AI returns a guideline answer with a source citation. No borrower data in the loop.

Credential management: Current state uses .env files (acceptable for single-operator development). Pre-deployment migration to AWS Secrets Manager or HashiCorp Vault with rotating credentials.

What Happens If Claude Goes Down

The income calculation engine does not use Claude. Deterministic rules engine. DTI, qualifying income, asset seasoning — all work without any AI API. An Anthropic outage does not affect income calculations.

• Frequently-asked guidelines are cached locally after first query. Common FHA questions return from cache without an API call.
• If Claude is unreachable: cached answers served immediately. Complex or uncached queries queued with estimated resolution time to LOs.
• Anthropic guarantees 99.9% uptime. Historical outages are rare and brief.

Model portability: RAG architecture is model-agnostic. Vector store is independent of inference provider. OpenAI or Google Gemini can be hot-swapped with a single configuration change. Guideline documents don't move. LO interface doesn't change.

2 AM Support — Who Picks Up the Phone

Current state — named and acknowledged: Nobody holds a pager today. Jasson is on-call. Alerted via Telegram, iMessage, and email on any service failure. Mean time to acknowledge: under 5 minutes during waking hours. Gaps during sleep hours are real and are acknowledged honestly.

90-day target:

• Cloud infrastructure with PM2 auto-restart handles the majority of failure modes without human intervention
• PagerDuty or equivalent for escalation that wakes a human when auto-restart fails
• Health check endpoints on every service, polled every 60 seconds
• Runbook documented for every known failure mode

At Rapid (if pilot proceeds): Kyle's team takes operational ownership for anything running in Rapid's environment. This is the natural end state for an internal tool. The CIO's infrastructure team runs the ops. Support SLA becomes Rapid's internal standard.

Subscription Tiers and Path to $50K/Month

Path to $50K/month (combined portfolio):

How This Works Practically — Two Tracks

Track 1 — Rapid Architecture Review (Kyle's gate role):

Track 2 — 9 Enterprises Advisory (relied upon, not just consulted):

AI Does Not Make Underwriting Decisions

• Source citations required: Every AI-generated guideline answer includes a source citation. The LO verifies against the source. The AI is a lookup accelerator, not an authority.
• Income calculations are deterministic: The rules engine is auditable, reproducible, and produces the same output for the same input every time. An underwriter can trace any income calculation step by step.
• Confidence flagging: Low-confidence answers are flagged for human underwriter review before the LO acts on them.
• Operations log: Daily log of all queries, answers, confidence scores, and flags.
• Architecture review: Kyle's role at Rapid is built into the deployment governance model permanently, not just during the pilot.

Honest Schedule, Not Optimistic Schedule