Ai Architecture — Investor Intelligence
12 Ai Systems. 3 Models.
One Intelligence Layer.
The only PropTech platform with cross-session semantic memory, Ai semantic caching, multi-model cost routing, and a proprietary SLM pathway. No competitor has more than 2 Ai systems. Most have zero.
The Ai Moat That Compounds Daily
Most Ai companies are resellers — they wrap an API call in a UI and charge a margin. When the API provider raises prices, their margins compress. When a competitor copies their prompt, they have no defense. RA's architecture is fundamentally different: semantic caching creates a data moat that grows with every user. Cost routing creates margin expansion that improves with scale. The SLM pathway creates a proprietary model no competitor can replicate. This is not an Ai feature. It is an Ai company.
Section 1 — Production Ai Stack
The 12 Ai Systems — All in Production
No competitor ships more than 2 Ai systems. Most ship zero. RA ships 12 in production.
1. Nelo Ai Concierge
Claude Opus 4.6 with native tool use, pgvector 1536-dim memory, 14 behavioral types, cross-session recall via embedding similarity, 6-portal intelligence. The core brain that orchestrates every other system.
2. Voice Ai Pipeline
Twilio → Deepgram nova-3 → Claude → ElevenLabs. Sub-300ms end-to-end latency. 10 languages with dedicated voice profiles. Replaces $60K/yr leasing labor per property.
3. LangGraph Orchestrator
9 specialized agents: LeadQualifier, TourScheduler, ApplicationReviewer, ScreeningOrchestrator, LeaseGenerator, InsuranceActivator, PaymentInitializer, AccessProvisioner, MoveInCoordinator. Full lead-to-lease autonomous pipeline.
4. FairScreen ISP
3-model consensus screening. Disparate impact analysis per decision. SHA-256 audit trail. 17 jurisdictions enforced. Patent pending. Any model disagreement routes to human review.
5. Z3 Policy Gate
Microsoft Z3 satisfiability solver. 6 verification domains. Mathematical proof of no disparate impact BEFORE execution — not after. Formal methods applied to housing policy for the first time in PropTech.
6. ZKP Verification (Groth16)
Zero-knowledge proof tenant qualification. Proves income eligibility without revealing salary. Proves credit threshold without exposing score. Patent-eligible with zero prior art in PropTech.
7. DAV Trust Fabric (ISO 18013-5)
Digital identity verification via mobile driver license standard. Government-grade authentication without exposing unnecessary personal data. Fraud prevention built into the transaction flow.
8. Ai Copilot
Listing descriptions, flyers, video tours, social media, email campaigns. Tri-LLM routed by task complexity. Subscription-tier aware — Pro gets efficient, Elite gets premium, Enterprise gets unlimited.
9. 3DGS Virtual Tours
Phone-captured 3D Gaussian Splatting. 4.2M splats per tour. Metric calibration for real-world measurements. Semantic layers for room-level intelligence. Forensic watermarks. Patent pending.
10. HAL Access Control
4-vendor hardware abstraction layer: ButterflyMX, Brivo, Allegion, Seam. One API controls all doors. Vendor swap in config, not code. Zero locksmith visits for access provisioning.
11. Ai Governance Engine
34-control governance framework. 28 implemented. Runtime enforcement, approval queues, output scanning, endpoint allowlisting. Every Ai action auditable, every decision reversible.
12. TokeniMax Cost Engine
Prompt compression + intelligent routing + semantic caching + budget governance. 84.5% compound COGS reduction pipeline. The system that turns Ai from a cost center into a margin engine.
Section 2 — Tri-LLM Engine
Three Brains. One Decision.
Think of Nelo like a consulting firm with three specialists. Need research? Gemini (cheapest). Need legal precision? Claude (most careful). Need a quick task? ChatGPT (fastest general-purpose). If one is unavailable, the others cover. No single point of failure.
Gemini 3.1
Research & Ops — 0.8x cheapest
Lead enrichment, Deep Research, multilingual intake, 13 artifact types, browser automation. The workhorse that handles 60%+ of all queries at the lowest cost.
Claude Opus 4.6
Premium Synthesis — 1.0x primary
Nelo core intelligence, investor memos, compliance analysis, safety-critical reasoning. The brain behind every decision that matters.
ChatGPT 5.4
Copilot & Orchestration — 1.5x fallback
Workflow orchestration, file tasks, automatic failover, circuit breaker. The reliable fallback that ensures 100% uptime.
Cheap-First Routing
Every request starts at the lowest-cost model that can handle it. Simple FAQ? GPT-5.4-mini at $0.15/MTok. Only when the task requires premium reasoning does Nelo escalate to Claude Opus 4.6 at $15/MTok. This is why margins improve with scale — most queries never need the expensive model.
Automatic Failover
Health checks every 30 seconds. Circuit breakers prevent cascading failures. If Anthropic is down, Gemini takes over. If Google is down, OpenAi covers. Three-provider redundancy means Nelo never goes offline.
Subscription-Aware
$89 Pro gets cheap-first routing. $149 Elite gets premium model access. Enterprise gets unlimited. Same Ai, different depth — controlled by entitlement config, not code changes.
Section 3 — Cross-Module Intelligence
One Ai. Every Vertical. Complete Context.
Nelo is not a chatbot bolted onto a dashboard. It is the intelligence layer woven through every module — with context that spans across all of them.
| Module | What Nelo Does | Real Example |
|---|---|---|
| Rental Listings | Writes descriptions, answers tenant questions 24/7, qualifies leads | Tenant asks "is parking included?" at 2am — Nelo answers instantly from property data |
| Property Management | Maintenance requests, lease renewals, payment processing by voice | Tenant calls about broken dishwasher — Nelo creates work order, notifies vendor, schedules repair |
| FairScreen | Orchestrates 3-model consensus, explains decisions, generates adverse action notices | Screening flags issue — Nelo drafts FCRA-compliant adverse action letter in 3 seconds |
| HOA / Glass HOA | Owner inquiries, board agendas, violation notices | Owner asks "why did assessment increase?" — Nelo pulls Z3-verified calculation and explains |
| Campus Guardian | CARA hour monitoring, coach alerts, audit reports | Coach approaches hour limit — Nelo alerts compliance officer with remaining allowance |
| Senior Housing | Medication reminders via Alexa, family wellness reports, resident voice queries | Resident asks "what's for dinner?" — Nelo responds with today's menu |
| Commercial RE | Lease abstraction, LL97 penalties, tenant financials | New lease uploaded — Nelo extracts 47 key terms in 30 seconds |
| Lead Marketplace | Enriches leads with Gemini, scores intent, routes to agents | New lead — Nelo generates 13-artifact intelligence package before agent sees it |
| Agent Services | Flyers, social posts, showing scheduling, lead follow-up | Agent says "create flyer for 432 Park" — Nelo generates branded PDF in 8 seconds |
| Voice Pipeline | Answers calls in 10 languages, qualifies callers, schedules showings | Spanish caller asks about 2BR — Nelo responds in Spanish, qualifies budget, books tour |
| Alexa Self-Tours | Property Q&A during in-unit visits, captures leads at peak interest | Visitor in kitchen asks countertop material — Nelo answers from property specs |
| Insurance Partners | Coverage recommendations during lease signing | At signing, Nelo presents Lemonade at $12/mo — 72% attach rate |
Section 4 — SLM Flywheel
Cost Trajectory — Ai That Gets Cheaper With Scale
| Stage | External LLM % | Cost/Query | Gross Margin | Mechanism |
|---|---|---|---|---|
| Year 0 (Now) | 100% | $0.003 | 80.7% | Full external API |
| Year 1 | 30% | $0.0008 | 88% | Semantic cache 70% |
| Year 2 | 10% | $0.0003 | 93% | Fine-tuned SLM |
| Year 3+ | 5% | $0.0001 | 95%+ | Proprietary PropTech SLM |
Data Collection (Now)
441 Prisma data models capture every interaction with structured metadata: intent classification, entity extraction, resolution outcome, satisfaction signal, response latency, model used, cost incurred. This is not logging — it is deliberate SLM training data architecture.
Semantic Cache Maturity (Year 1)
Cache hit rates climb 30% → 50% → 70%. Nelo handles the majority of queries without any external LLM call. Cached responses become the initial training corpus — proven correct answers to real PropTech questions.
Fine-Tuned SLM (Year 2)
Millions of cached Q&A pairs fine-tune a 7B-13B parameter model specifically for PropTech. Handles 90% of routine queries at near-zero marginal cost. Claude and Gemini reserved for complex reasoning only.
Proprietary PropTech SLM (Year 3+)
Fully trained RA SLM understands real estate at a depth no general-purpose model can match. Lease structures, FARE Act, HOA bylaws, CARA rules, HIPAA constraints, commercial escalation patterns. External LLM usage drops to 5%. Gross margin exceeds 95%. 3-year head start no competitor can replicate.
Section 5 — Training Data Architecture
Every Conversation Today Is Training Data for Tomorrow's Proprietary Model
441 Prisma data models capture every interaction with structured metadata: intent classification, entity extraction, resolution outcome, satisfaction signal, response latency, model used, cost incurred. This is not logging — it is deliberate SLM training data architecture. By Year 3, RA will have millions of real PropTech Q&A pairs that no competitor can access.
Continuum Memory (459 LoC)
pgvector 1536-dim embeddings store every interaction. 14 memory types for user preferences, property interests, communication style, search patterns, behavior patterns, domain knowledge, feedback — plus fast memory types for session context. Nelo remembers you 3 months later.
Semantic Cache (628 LoC)
pgvector-backed query/response cache with cosine similarity matching. When tenant B asks the same question tenant A asked, the cached answer returns instantly — no LLM call. 70% hit rate at maturity. Each new user warms the cache for future users. Network effect.
Cost Routing (3,388 LoC)
Every task type maps to cheap, standard, or premium candidates. Budget ledger tracks per-model costs. Token governor enforces input/output limits per quality tier. The economics engine behind the 97% cost reduction.
Subscription Entitlements (281 LoC)
5 subscription tiers with per-feature Ai budgets, premium escalation quotas, and feature-level overrides. Ai quality scales with subscription price — controlled by config, not code.
Section 6 — Competitive Intelligence
The Gap Is Not a Feature Gap. It Is an Architecture Gap.
Every competitor named below was evaluated against RA's 12-system Ai stack. The results are definitive.
| Ai Capability | RealRiches (RA) | ||||||
|---|---|---|---|---|---|---|---|
| Ai Systems Count | 12 production | 2 (leasing + voice) | 1 (pricing algo — DOJ flagged) | 0 | 1 (Realm-X copilot) | 0 (SkyTour is Matterport) | 0 |
| Cross-Session Memory | 14 types, pgvector 1536-dim | Session-only | None | None | None | None | None |
| Semantic Cache | 70% hit rate at maturity | None | None | None | None | None | None |
| Multi-Model Routing | 3 providers + cost tracking | Single model | Single model (DOJ flagged) | None | Single model | None | None |
| Voice Ai Pipeline | Sub-300ms, 10 langs | Proprietary NLP (pre-LLM, $392M to build) | None | None | None | None | None |
| Formal Verification | Z3 solver, 6 domains | None | None | None | None | None | None |
| Zero-Knowledge Proofs | Groth16, patent-eligible | None | None | None | None | None | None |
| Agentic Orchestration | LangGraph 9-agent pipeline | Intent classifiers (not agentic) | None | None | None | None | None |
| 3DGS Virtual Tours | Phone-captured, patent-pending | None | None | None | None | SkyTour (acquired | None |
| Smart Access (HAL) | 4-vendor abstraction | None | None | Basic keycards | None | ShowTime (basic) | Basic |
| Ai Governance | 34 controls, runtime enforcement | None | Under DOJ consent decree | None | None | None | None |
| SLM Training Pathway | 441 Prisma models, JSONL export | Locked to proprietary NLP | None | None | None | None | None |
| FARE Act Compliance | Built-in, fail-closed, 17 jurisdictions | None | Under DOJ investigation | None | None | Under FTC suit | None |
| LLM Price Resilience | 85% absorbed via cache | 100% pass-through | 100% pass-through | N/A | 100% pass-through | N/A | N/A |
| Cost to Replicate | $12-25M + 36-48 months | — | — | — | — | — | — |
EliseAi raised $2.25 billion dollars to build TWO Ai systems — leasing automation and voice. RealRiches ships TWELVE. Yardi's codebase is 42 years old — it was written before the internet existed. AppFolio's Realm-X is a single copilot with no memory, no caching, no formal verification. RealPage is under DOJ consent decree for the only Ai system they built. Zillow spent $1.2 billion acquiring Matterport and still doesn't have Ai screening, voice, or smart access. The gap is not closing. It is widening.
Section 7 — TokeniMax Enhancement
How TokeniMax Supercharges Nelo's Economics
| Tier | Price | COGS Before TokeniMax | COGS After | Margin After |
|---|---|---|---|---|
| Pro | $89 | $14.23 | $2.14 | 97.6% |
| Elite | $149 | $32.26 | $4.85 | 96.7% |
| Team | $499 | $89.95 | $13.49 | 97.3% |
| Enterprise | $1,299 | $263.61 | $39.54 | 96.9% |
Compound Savings Pipeline
3,000 requests/subscriber → 70% cached (free) → 80% routed to cheap models → 67% fewer tokens via prompt compression → Net: $32.26 → $4.85 per subscriber (85% reduction)
See the Full Ai Architecture in Action
Schedule a live walkthrough with the founder. Every system described above is running on production code — 2.02M lines of TypeScript.