LOUIE.BRAIN v2.4 PATTERNS 322+ AGENTS ONLINE 29 MODEL Local LLM TOOL-USE Claude Tool Loop INFERENCE On-Device CLOSED-LOOP ACTIVE NETWORK EFFECT BUILDING LOUIE.BRAIN v2.4 PATTERNS 322+ AGENTS ONLINE 29 MODEL Local LLM TOOL-USE Claude Tool Loop INFERENCE On-Device CLOSED-LOOP ACTIVE NETWORK EFFECT BUILDING
AI Platform Deep Dive

The On-Device Brain
That Runs Louie Auto Group

Not a chatbot you rent. A reasoning engine you own — running locally, learning from every deal closed, getting smarter every night.

Ollama Local Inference
qwen2.5:3b + 8b
Claude Tool-Use Layer
29 Active Agents
322+ Learned Patterns
Active Agents
29
across all departments
Brain Patterns
322+
learned from real outcomes
Claude Tools
12
tool-use loop, not chat
Inference Mode
On-Device
zero latency, zero API cost
Nightly Learning
1:30 AM
cron sync + reweight
Section 01

The Brain Architecture

Three-layer stack: on-device fast inference, rule-based routing, and Claude API for generative tasks. Each layer does exactly what it's best at.

Layer 1 — On-Device Inference
Ollama Runtime
Local HTTP inference server. Zero API cost, zero latency, zero data egress. Runs on dealership hardware or cloud VM.
↓
qwen2.5:3b — Fast Decisions
Sub-100ms responses for scoring, routing, and real-time deal guidance. Lives in RAM, never sleeps.
↓
qwen2.5:8b — Deep Reasoning
Heavier model for compliance analysis, risk assessment, and complex deal structuring where accuracy beats speed.
Layer 2 — Rule Router
Pattern Match Engine
322+ learned patterns from real closed deals. High-confidence patterns skip LLM entirely — deterministic and fast.
↓
Confidence Threshold
Pattern confidence < 0.72 → escalate to local LLM. < 0.55 → escalate to Claude API for generative reasoning.
↓
Closed-Loop Scorer
Every decision logged against outcome. Funded = reinforce. Declined = penalize. Weights recomputed nightly at 1:30 AM.
↓
Lender Weight Cache
getWeights(dealershipId) — live per-dealer lender score map consumed by matchLenders(). Isolated per rooftop.
Layer 3 — Claude API (Generative)
Claude AI
Handles tasks that require genuine language generation: BDC scripts, compliance narratives, deal summaries, market briefs.
↓
Tool-Use Loop
Claude doesn't just answer — it reasons, calls tools (inventory, lender matrix, compliance), acts, then debriefs. Real agentic loop.
↓
Outcome Feedback Bus
Every AI action — agent run, deal routed, script sent — logs to intelligence_log. Feeds nightly brain reweight.
↓
WAL-Mode SQLite (per module)
Each module owns its database. busy_timeout=5000, WAL mode, zero shared-lock contention across 4 PM2 workers.
Decision Routing Logic
confidence ≥ 0.72 → Pattern match (deterministic) <5ms · no model call
0.55 ≤ confidence < 0.72 → Ollama local LLM 30–150ms · on-device
confidence < 0.55 → Claude Tool Loop (12 tools active) 1–4s · generative
compliance / legal → qwen2.5:8b + rule overlay deterministic + AI verify
// brain-agent.js — real tool-use loop const response = await anthropic.messages.create({ model: 'claude-sonnet', tools: brainTools, // 12 tools: inventory, lenders, compliance… messages: [{ role: 'user', content: dealerGoal }], max_tokens: 4096, }); // Streams tool_use blocks → executes → loops until stop_reason='end_turn' while (response.stop_reason === 'tool_use') { const results = await executeTools(response.content); // append tool_result blocks, re-call — real agentic loop }
Section 02

29 AI Agents

Every department has dedicated intelligence. Not a single chatbot that handles everything — purpose-built agents wired to the data they need.

01
Deal Router
Scores incoming deals against 322+ patterns. Routes to the optimal lender in under 5ms. Penalizes declined lenders, reinforces funded ones nightly.
Core Loop
02
BDC Script Generator
Generates live, personalized outbound scripts from real CRM data — equity position, payment history, service visits. Not templates. Actual context.
Claude API
03
Louie Voice Agent
AI receptionist + outbound BDC caller. Works unsold, equity, and service lists 24/7. Answers inbound cold. Plugs into Twilio/Telnyx via provider layer.
Live
04
Ad Engine
Builds Google PMax and Meta Advantage+ campaigns from live inventory and real CDP audiences. Pushes to ad accounts via API when keys are connected.
Live
05
Investment Scorer
Inventory profit-scoring, built in — scores every unit on days supply, market demand, reconditioning cost, and gross potential. Priority queue for the lot manager.
Live
06
Risk Assessor
Evaluates deal risk across 14 dimensions: LTV, PTI, employment stability, credit tier, co-buyer exposure, state rate caps. Flags before funding.
Core Loop
07
Compliance Checker
Real-time check against FTC Safeguards, TILA, UDAP, state rate caps (all 50 states + DC), Red Flags Rule. Flags violations before the deal prints.
Live
08
Credit Trend Analyzer
Tracks customer credit tier migration. Flags customers who crossed from subprime to near-prime — your next upgrade opportunity. Real pattern, not synthetic.
Live
09
Equity Alert Engine
Monitors the entire customer portfolio for positive equity events — market price moves, payoff milestones. Triggers BDC outreach automatically.
Live
10
Morning Briefing Generator
7 AM daily briefing built from real data: aged units, hot leads, lender rate changes, FRED economic signals, overnight deal outcomes. Not a dashboard — a briefing.
Claude API
11
Close Coach
Live deal coaching inside the F&I desk. Watches deal structure in real time, flags objection patterns, suggests counter-offers from closed-deal history.
Claude API
12
Autonomous Queue Manager
Prioritizes the BDC queue by predicted close probability. High-equity + recent service visitor + credit tier upgrade = top of list. Dynamic, not static.
Core Loop
13
Lender Match Agent
matchLenders() consumes live weight map. Tier-1 to tier-5 tiers with 42 lender matrix. Weights from real funded/declined outcomes, recomputed nightly.
Core Loop
14
OFAC Screening Agent
Live Treasury SDN list check on every customer write. Not a mock — real API call to OFAC. Blocks deal submission on match. Audit log for every check.
Live
15
Desk Four-Square Coach
Watches the four-square negotiation live. Calculates gross on every move, suggests counter-offer strategy from won/lost deal history.
Claude API
16
Rate Watch Agent
Pulls FRED UMCSENT, prime rate, and Fitch lender data. Briefs the F&I manager when rates shift. Live signals, not hardcoded numbers.
Live
17
Negative Equity Coach
Structures negative equity roll-ins to minimize lender exposure while protecting gross. Knows which lenders tolerate how much negative at each tier.
Core Loop
18
Service-to-Sales Bridge
Identifies service customers within 90 days of lease end or positive equity. Routes to BDC as high-intent prospects before they shop a competitor.
Live
19
Market Days Supply Agent
Compares current inventory mix vs local market days supply via MarketCheck API. Surfaces over-aged segments and under-stocked opportunities.
Live
20
Book Value Intelligence
Real MarketCheck auction data + NADA/KBB fallback for trade valuations. Prevents overpays on trade-ins. Sourced from live auction lanes, not static tables.
Live
21
Simulation Engine
Runs synthetic deals through the lender matrix to calibrate weights before going live. 1.34M+ simulated declines feed the credit tier model.
Core Loop
22
BHPH Collections Agent
Buy-here-pay-here payment prediction: likelihood to pay, ideal contact time, escalation triggers. Real amortization engine with TILA-compliant output.
Live
23
Knowledge Retention Agent
Captures winning patterns from top performers. When your best F&I manager leaves, his close patterns stay in the brain. Year 3 smarter than Year 1.
Core Loop
24
Onboarding Intelligence Agent
Guides new dealers through integration setup, initial brain calibration, and first-deal validation. Reduces onboarding from weeks to hours.
Live
25
Campaign Performance Scorer
Scores ad campaign performance against actual vehicle sales. Closes the loop from ad spend to front/back gross. Attribution on data you own.
Live
26
Competitor Intel Watcher
Monitors competitor pricing and inventory levels via MarketCheck. Surfaces when you're priced out of market on specific segments. Automated weekly report.
Sim Mode
27
GL Integration Agent
Posts double-entry journal entries to GL on every funded deal. Debit contracts-in-transit, credit floorplan. Real accounting logic, not a summary export.
Live
28
FTC Safeguards Tracker
Monitors completion of all 9 FTC Safeguards rule requirements. TOTP MFA enforced. Security training log. Derives from live ftc-safeguards.db — not hardcoded.
Live
29
Nightly Reweight Cron
1:30 AM: sync simulation outcomes → closed-loop DB → recompute lender weights → update brain patterns → purge stale cache. The brain that grows while you sleep.
Core Loop
Section 03

The Learning Engine

322+ live patterns extracted from real deal outcomes. Not synthetic training data — actual funded and declined deals from the dealership floor.

Pattern Category Patterns Confidence
Lender Routing — Prime 48
94%
Lender Routing — Subprime 41
89%
Equity Trigger Points 37
91%
BDC Outreach Timing 29
82%
Deal Structure (F&I) 34
87%
Negative Equity Tolerance 22
76%
Credit Tier Migration 31
88%
Compliance Risk Flags 44
96%
Inventory Pricing Signals 26
79%
Service-to-Sales Conversion 10
71%
How Patterns Are Born
01Deal submitted → lender decision logged to lender_outcomes
021:30 AM cron extracts 30-day outcome clusters
03High-frequency winning combos become named patterns
04Patterns assigned confidence score; low-confidence → LLM verify
05Next day, 5ms pattern match before any model call
Year 3 vs Year 1
Every rep who closes a deal trains the brain. Every rep who leaves — their patterns stay. Year 3 with 900 closed deals is exponentially smarter than Year 1 with 50. That's the moat you can't rent. You own it.
AI Decision Confidence by Department
Average confidence score across agent decisions
Brain Learning Curve
Patterns accumulated over deployment lifetime
Agent Workload Distribution
Share of total agent decisions by category
Section 04

Claude Tool-Use: 12 Active Tools

Claude doesn't answer from memory. It reasons, calls tools against live data, acts on the result, and debriefs. This is a real agentic loop — not a chatbot with car knowledge.

📦
inventory_analysis
Queries live Vehicles table. Filters by make/model/days/price. Returns aged units, market days supply delta, and reconditioning cost exposure.
🏦
lender_routing
Submits deal parameters to matchLenders() with live weight map. Returns ranked lender list with estimated approval probability per tier.
📐
deal_structure
Calculates payment permutations across term/rate/down combinations. Checks against PTI and LTV limits. Outputs compliant deal structures.
⚠️
risk_scoring
Scores deal risk across 14 dimensions. Pulls live lender_outcomes to weight state and customer-segment risk. Returns 0–100 risk score with flags.
⚖️
compliance_check
Validates against all 50-state APR caps, TILA disclosure requirements, Red Flags Rule, UDAP statutes. Returns pass/fail per rule with remediation note.
📣
ad_campaign_builder
Builds Google PMax / Meta Advantage+ campaign specs from live inventory + CDP audiences. Returns campaign JSON ready to push via ad provider API.
📞
bdc_campaigner
Pulls CRM segment (equity, unsold, service due). Generates personalized outreach scripts. Routes to voice/SMS via Twilio/Telnyx provider layer.
📊
market_intelligence
Pulls FRED UMCSENT, prime rate, Fitch lender data. Returns current economic context for deal structuring and rate watch briefings.
📈
credit_trends
Analyzes credit tier migration across the customer portfolio. Surfaces customers who crossed tier boundaries in the last 90 days. Live data.
💎
investment_scoring
Runs profit-tier inventory scoring on every unit, the kind standalone pricing tools charge extra for: demand velocity, margin potential, reconditioning delta, days-to-market risk.
📅
market_days_supply
MarketCheck live API — compares current stock mix vs local market turn rates by segment. Flags over-aged and under-stocked pockets.
🏆
lender_scorecard
Returns per-lender performance stats from closed-loop outcomes: approval rate, funded rate, avg gross, avg days to fund. Live 30/60/90 day windows.
Why Tool-Use Matters
A chatbot answers from its training weights. A tool-use agent answers from your actual data — this week's inventory, last month's lender outcomes, your customer's real equity position. Same model. Completely different answers. That gap is the product.
The Reasoning Chain
User: "Find equity customers for a spring push"
Louie: → calls credit_trends
→ calls inventory_analysis
→ calls bdc_campaigner
→ calls ad_campaign_builder
Done: Campaigns built + scripts queued
Section 05

The Network Effect

When 100 dealers share anonymized outcomes, every dealer's brain gets smarter. This is a moat neither a legacy DMS vendor nor a data activation platform can buy or replicate.

1
Dealer network — growing
322 Shared Patterns
42 Lenders Mapped
100x Network Multiplier at 100 Dealers
  • 🔒
    Anonymized by designNo dealer sees another dealer's customers, prices, or deal details. Only outcome patterns are shared — win/lose signals, not PII.
  • 📡
    Peer benchmarksYour close rate vs. comparable dealers. Your lender mix vs. region. Surfaced in the Morning Briefing automatically.
  • 🏦
    Lender leaderboardWhich lenders are funding at the best gross for your segment, right now — across the network. Not from a rep's pitch deck.
  • 🚨
    Fraud WatchStraw purchase pattern seen at Dealer A → all dealers in the network are protected within 24 hours of the nightly sync.
  • 📈
    The data moatThe incumbent DMS vendors control roughly 80% of the DMS market but their data is siloed. LouieAuto network data is shared intelligence. No legacy DMS vendor can buy this — they'd have to ask their own dealers to give it back.
  • 🌐
    BHPH to franchiseBuy-here-pay-here dealers, independents, and franchise stores all learning together. Payment performance from BHPH makes the credit model smarter for everyone.
1
Dealer → 322 patterns
baseline brain
→
10
Dealers → 4,000+ patterns
regional intelligence
→
100
Dealers → 40,000+ patterns
uncopyable moat
Legacy DMS moat is lock-in. LouieAuto's moat is collective intelligence. One is a trap. The other is a compounding asset.
For Acquirers & Strategic Partners

The AI Stack Is Built.
It Ships With the Platform.

29 agents. 322+ patterns. 12 tools. Closed-loop learning. On-device inference. The moat that gets deeper every night.

View Acquisition Brief Technical Diligence FAQ Watch Brain Simulations Live
v3.5.3
Report Issue