Louie's recommendations draw from six categories of dealership data, all stored locally on your server. No data is sent to a third-party training pipeline. No customer records leave your network.
- Inventory Live stocked vehicles — VIN, year, make, model, trim, mileage, cost, days-on-lot, auction source, condition, and reconditioning cost. Used for pricing recommendations and aging alerts.
- Deal History Every funded and penciled deal: vehicle, customer credit tier, lender, rate, term, reserve, F&I product penetration, and deal outcome. Used for lender routing and PVR coaching.
- Lender Outcomes Approval / decline / counter results from your submitted deals, organized by lender, FICO band, LTV, and vehicle type. Weights auto-adjust nightly via closed-loop learning.
- Customer Interactions CRM touchpoints: lead source, visit history, text/email exchanges, and appointment outcomes. Used for BDC prioritization and follow-up timing. No interaction content is used in AI prompts.
- Market Signals — FRED UMCSENT University of Michigan Consumer Sentiment Index fetched live from the Federal Reserve Economic Data API. Used to contextualize market timing recommendations.
- Market Signals — MarketCheck Auction Real-time auction price data for comparable vehicles via MarketCheck API. Used in book-value and trade appraisal calculations to anchor to live market, not stale guides. Optional add-on: $299/month direct from MarketCheck — not billed through Louie. Manage at my.marketcheck.com.
Louie's core intelligence runs locally, on your server, via an on-device Ollama model — nothing about your dealership's data leaves your machine for a routine recommendation. When the local model can't confidently handle a request, it falls back to Anthropic's Claudeclaude-sonnet-4-6 as a structured reasoning engine, still running inside your server process. Neither model retains context between sessions, and neither has access to external systems beyond what is explicitly listed in Section 01.
Each request assembles a context packet before the model is called:
- Store snapshot: current inventory counts, active deal pipeline, and pending follow-ups
- Role-specific filter: salesperson sees their deals; manager sees store-wide; owner sees all
- Real-time signals: latest FRED UMCSENT reading + MarketCheck comparable pricing (if relevant to the query)
- Historical context: trailing 90-day lender approval rates and F&I product penetration for this store
- User query: the specific question or task submitted by the dealer employee
Three categories of action require explicit human approval before Louie can proceed. These gates cannot be bypassed through any API call or automation rule.
LouieAuto runs on a private VPS you own or lease. There is no multi-tenant cloud instance. No other dealership's data shares your database, network, or process space.
-
Private VPS — not multi-tenantYour server, your instance. No shared infrastructure with other dealers. Data stays inside your network perimeter at all times.
-
AES-256-GCM encryption at restAll PII, SSN, and financial data is encrypted at rest via LUKS full-disk encryption on the dealership server. The encryption key is stored separately from the data volume.
-
JWT-gated API routesEvery API endpoint requires a valid JWT. Tokens carry dealership ID and role claims. Requests without valid tokens return 401. No public data endpoints expose customer records.
-
Role-based data visibilitySalesperson can see their own deals and customers. Finance manager sees store-wide deal pipeline. Owner sees all modules including owner board and download analytics. No role can escalate its own permissions.
-
No third-party analytics on financial dataNo Google Analytics, Segment, Mixpanel, or any third-party pixel has access to customer financial records, deal data, or credit information. External analytics (if any) receive only page-view data.
Every recommendation Louie generates is written to an immutable audit log at the time of generation. The log entry captures what data signals were used, what confidence tier was assigned, and what the human ultimately did with the recommendation.
{
"action": "lender_route",
"module": "f-and-i",
"signals": ["fico_588", "ltv_112", "term_72", "lender_history_westlake"],
"confidence": "high",
"recommendation": "Route to Westlake Financial — 84% approval probability",
"human_outcome": "accepted",
"ts": "2026-06-11T14:23:07.881Z",
"user_role": "finance_manager",
"dealership_id": "default"
}
All recommendation logs are permanently retained in your local database and included in encrypted backups. They are never deleted by system processes.
Retention periods follow IRS, CFPB, and state regulatory requirements. All data is stored locally in your encrypted database. You can export any category at any time.
| Data Category | Retention Period | Regulatory Basis |
|---|---|---|
| Deal records | 7 years | IRS Schedule C / state audit requirements |
| Customer PII | Until deletion request | FCRA / state privacy law (CCPA, etc.) |
| AI recommendation logs | 3 years | FTC Safeguards Rule audit documentation |
| Credit consent records | 25 months | ECOA / Reg B adverse action window |
| Call and text transcripts | 3 years | TCPA + state recording disclosure laws |
| Encrypted backups | 90-day rolling local | FTC Safeguards backup requirement |