← Back to Interaction Rule Set
Product Requirements Document V5
AI-Powered Mortgage Education Platform · Version 5.1
Abstract
PreFi's Clarity Engine is an AI-powered mortgage education platform that empowers homeowners to explore refinancing options through neutral, visual scenario comparisons. Unlike lead-generation platforms that sell user data to multiple lenders, PreFi presents 3-4 scenario cards per user intent with clear trade-offs, interactive calculators, and beautiful visualizations, without steering toward any single option.
The Clarity Engine is an intelligent financial guide that helps homeowners understand their refinancing options through personalized scenarios, visual explanations, and tradeoff exploration, without requiring immediate verification or commitment.
In its Alpha form, the Clarity Engine prioritizes clarity before conversion: delivering meaningful insights using minimal, self-disclosed inputs, and progressively deepening accuracy only when the user opts in. The experience is designed to feel advisory, unbiased, and empowering, not transactional or sales-driven.
Core Problem
The mortgage industry optimizes for lead volume, not consumer clarity. LendingTree sells data to 5-10 lenders who compete with aggressive calls. Bankrate optimizes for clicks. Consumers under 40 years old (70%+) ask friends on social media because they don't trust industry sources.
Alpha Solution
Stage 1 anonymous exploration where AI discovers user intent, educates with 5th-grade-level language and visualizations, and presents scenario cards showing all viable paths with transparent trade-offs. User controls exploration. No steering, no recommendations, no pressure.
Future Vision
Stage 2 adds verification (credit/income/assets) to show scenarios with real numbers. Stage 3 adds a lender marketplace where the user selects a path, then chooses a lender. The platform operates on a freemium model: free for consumers, lenders pay on lead fulfillment.
The Catalyst
PreFi Clarity Engine is the catalyst for building world-class AI-powered lending infrastructure. The emotional intelligence, scenario generation, and platform architecture we build for refinance become the foundation for future capabilities.
But first: Get PreFi right. Prove the model. Win with refinance.
1. Business Objectives
1.1 Primary Objectives
- Establish trust as a neutral educator: Become the go-to mortgage education platform for consumers who value transparency over being sold to
- Prove empowerment model: Validate that users prefer exploring multiple scenarios over receiving a single AI recommendation
- Build defensible moat: Superior UX + neutrality + emotional intelligence + multi-modal learning (text + visuals + calculators) creates differentiation
- Enable platform scalability: Architecture supports future white-label deployments for mortgage brokers and financial advisors
- Validate stage-gated approach: Prove Stage 1 (validate) creates enough value to justify Stage 2 (verify) and Stage 3 (fulfill) investment
- Create foundation for future: Build the emotional intelligence and data infrastructure that enables broader lending innovation
1.2 The Fight We're Picking
We're redefining the mortgage lead market by putting consumer empowerment over lead volume.
We're picking a fight with how consumers currently get mortgage advice: lead generation (sell confusion as choice), loan officers (order-takers), and generic calculators (no context). We win by delivering expert optimization with radical clarity.
2. Emotional Intelligence Architecture
What separates PreFi from competitors is our 4-component emotional intelligence system. While Google provides static results and ChatGPT lacks the ability to validate borrowers' qualifications, confirm if borrowers can qualify for prescribed offers, and fulfill recommendation meets you where you are, guides your journey, supports how you feel, and shows what you need.
2.1 The Critical Gap
Research shows 40% of users get confused or frustrated during mortgage exploration. Without emotional intelligence, these users abandon mortgage applications.
The Gap: When someone asks about amortization three times, traditional systems keep explaining the same way.
Result: User abandons.
With Emotional Intelligence: System detects repeated confusion. Simplifies explanation. Offers visual alternative. User breakthrough.
2.2 Four-Component Architecture
Think of PreFi as a complete organism, not just a tool:
Oracle (Brain)
Purpose: Understands WHO the user is and predicts which scenarios will resonate with their archetype.
Functions:
- Detects user archetype: Analytical, Emotional, Practical, or Cautious
- Predicts which scenarios will resonate based on the archetype
- Adapts journey complexity and pacing to match user style
- Learns from 1M+ MonsterLead conversations (data moat)
Navigator (Spine)
Purpose: Orchestrates the journey in real-time, providing clear "next step" guidance and adapts in real time to changes in data.
Functions:
- Manages conversation flow across 3 stages
- Provides a clear "what happens next" at each step
- Auto-adjusts when the user hits roadblocks
- Maintains context across session interruptions
Advisor (Heart): The Critical Missing Piece
Purpose: Detects emotional state and adapts communication to make users feel understood, not alone.
Functions:
- Detects emotional state: confused, frustrated, confident, anxious, excited, overwhelmed, disengaged
- Adapts tone and communication style based on emotion
- Intervenes during crisis moments (simplifies, reassures, celebrates)
- Creates "PreFi actually got me" moments that drive word-of-mouth
Storyboard (Skin)
Purpose: Visual clarity layer that makes complex financial concepts accessible.
Functions:
- Scenario card comparisons with clear trade-offs that meet their personal goals
- Interactive calculators responding instantly (<200ms)
- Beautiful visualizations explaining concepts at 5th-grade level
- Mobile-first design with 60fps interactions for page transition speeds
2.3 Clarity Engine Behavioral Model by Stage
Stage 1: Anonymous Guidance (Alpha Required)
- Users can explore refinancing options anonymously with self-disclosed inputs only
- No account, no credit pull, no identity verification required
- Engine generates initial recommendations quickly using minimal inputs (e.g., balance, rate estimate, goals)
- Visual scenarios appear early and refine dynamically as users adjust inputs
- All calculations in Stage 1 are based on self-disclosed inputs
- A single-bureau soft credit pull is permitted only after user opt-in
- Prior to a soft credit pull, users must be shown a plain-language explanation of what will change (and what will not), and must explicitly opt in via a distinct action (not inline continuation).
- No hard credit pulls, income verification, asset verification, or lender integrations are allowed in Alpha
- All outputs clearly disclose assumptions and data limitations
- In Alpha, all behavioral adaptation occurs within a single session only. Cross-session learning, longitudinal memory, and proactive inference are explicitly deferred to Stage 2+
Stage 2: Saved & Enhanced Guidance (Post-Alpha)
- Users may opt in to save progress via lightweight authentication (email-first)
- Soft credit pull (single bureau) is permitted to improve constraint accuracy
- The engine remembers how users interacted with scenarios (e.g., sliders, tradeoffs explored) to infer preferences
- Reports and scenarios become shareable and downloadable
Explicitly Out of Scope for Alpha
- Hard credit pulls
- Lender offers or fulfillment
- Proactive "monitoring" or alerting about future refinance opportunities
2.4 Why This Creates Defensibility
After 10,000 users, PreFi's emotional intelligence becomes impossible to replicate:
- Data Moat: Emotional journeys compound over time. We learn which interventions work for which archetypes in which situations.
- Network Effects: More users leads to better emotional intelligence leads to higher completion leads to more users
- Category Positioning: First platform that emotionally understands financial decisions
- Word-of-Mouth: "PreFi actually got me" creates organic growth that competitors can't buy
3. The Magical Moment
Every transformative product has ONE interaction that makes someone immediately text their friend. For PreFi, it's the Crisis Intervention Moment, when the platform literally morphs to help you understand.
3.1 The Interaction
Trigger: User asks about the same concept (e.g., amortization) more than once, or expresses frustration ("I don't get it", "This is confusing", "Ugh", "I just want to know what this means for me and my scenario")
What Happens: The Advisor (Heart) detects a crisis. Instead of repeating the same explanation, the
ENTIRE INTERFACE GENTLY TRANSFORMS:
- Screen Transition (500ms gentle fade): Current view softly blurs
- AI Message Appears: "Let me try explaining this differently" (warm, reassuring tone)
- Visual Morphs: Complex chart animates into simpler version. Fewer labels, bigger numbers, clearer colors
- Alternative Explanation: Same concept, completely different analogy (e.g., "Think of it like paying off a car loan...")
- Interactive Element: "Want to see this as a video instead?" button appears
- Story Alignment: "Let's explain this in the context of your scenario" and the system will start using examples that pertain to what we know about the user versus generic context with an understanding not only of the "why," but the "why not"
3.2 Technical Specification
Animation Details:
- Duration: 500ms fade (smooth, not jarring)
- Easing: cubic-bezier (0.4, 0.0, 0.2, 1), "ease-in-out"
- Chart Simplification: Reduce data points by 60%, increase font sizes by 150%, change to high-contrast colors
- Micro-copy: Short (max 15 words), warm tone, second-person ("you"), active voice
3.3 Success Metrics
The Magical Moment works when:
- 75%+ of users who trigger intervention continue (vs abandon)
- 60%+ report "PreFi understood when I was stuck" in post-session survey
- Users mention it unprompted: "The way it helped me when I was confused was amazing" or "PreFi helped me identify the options that work for me and I felt confident selecting the right path forward"
This is the "holy sh!t" moment. This is what they text their friends about.
4. Emotional Design Language
Every pixel, every animation, every word choice is intentional. PreFi's design language creates "Calm Confidence", the feeling that complex financial decisions become clear, and you're supported throughout.
4.1 Primary Emotion: Calm Confidence
Not "Excited Energy" (too aggressive). Not "Clinical Precision" (too cold).
Calm Confidence feels like:
- A trusted advisor who's done this 1,000 times and knows exactly what to do
- Clear explanations that make you feel smart, not talked down to
- Deep understanding and intelligence around not only the "why" but the "why not"
- Steady, unhurried pace that says "Take your time, we're not going anywhere"
4.2 Color Psychology
Anti-Patterns (Never Use):
- Bright Red: Creates anxiety, pressure
- Neon/Saturated Colors: Feels cheap, untrustworthy
4.3 Animation Timing
Speed conveys emotion. PreFi deliberately moves SLOWER than typical apps.
4.4 Typography & Voice
Font Choices:
- Headings: Inter (modern, clear, trustworthy)
- Body Text: Inter (consistent family, easy to read)
- Numbers: SF Mono / Roboto Mono (distinct, precise)
4.5 Scenario Card Personality
Scenario cards need PERSONALITY. They're not "Scenario A, B, C". They're paths with names, icons, and emotional resonance.
5. User Journey with Emotional Intelligence
The Clarity Engine prioritizes early value delivery. Initial recommendations and visual scenarios are presented after a small number of inputs, then refined progressively as the user chooses to go deeper.
Users may explore lightly or engage in deeper scenario modeling; the system adapts without forcing a single linear interview path.
5.1 Primary Persona: Alex (30-year-old professional)
Context:
- Age 28-35, household income $100K-$200K, considering refinance or exploring ways to cut costs for monthly savings
- Mobile-native, values design quality, zero tolerance for poor UX
- Intimidated by financial jargon but highly educated
- Distrusts loan officers (feels like being sold to) or doesn't know enough about them or the profession to know that they're really just being sold to and that they might have their own motivations
- Wants smart decisions with confidence, at their own pace, not hour-long sales calls
5.2 Current Behavior (Before PreFi)
- Searches "why does my mortgage payment keep going up I thought it was fixed" on Google
- Searches "should I refinance" on Google
- Lands on Bankrate or NerdWallet, gets overwhelmed by options
- Asks friends on Facebook: "Has anyone refinanced lately?"
- Gets conflicting advice, remains confused
5.3 Journey with PreFi
1. Discovery
Friend says: "Try PreFi, it actually explains my scenario as it relates to mortgages without trying to sell you on anything so you can figure out the best path forward to reach your goals"
2. Entry
Visits PreFi website, sees: "Explore your mortgage options. No pressure, no steering." Clicks "Start Exploration"
Oracle: Begins detecting user signals
3. Example Conversation & Intent Discovery
AI asks warm questions to understand intent: "What brings you here today?"
Alex types: "Thinking about refinancing but not sure if it makes sense"
Oracle: Detects cautious language, thoughtful approach. Likely Cautious archetype
Advisor: Detects anxious (0.6). Adapts tone to be reassuring
AI responds: "That's a really common place to be. Mortgages can feel overwhelming. Let me help you figure this out, no pressure. What would a successful refinance accomplish for you?"
Alex: "I guess I want to lower my monthly payment? We're planning for a baby and could use the extra cash flow."
Oracle: Real intent = free up cash flow for life goal (baby)
Navigator: Prioritize lower payment scenarios after getting more information about current situation
4. Education
AI shows a visualization explaining amortization and compound interest at a 5th-grade level. Alex feels smart, not talked down to with the hyper personalized recommendation.
Advisor: Detects confidence increasing (0.7). User ready for scenarios
5. Scenario Cards with Soft Guidance
AI presents 3 paths with soft guidance:
"The Freedom Path addresses your priority (lower payment), here's why, and here are the trade-offs:"
The Freedom Path (Extend Loan Term)
- Save $450/month
- More total interest over life of loan
- Frees up cash flow immediately for baby expenses
The Fast Track (Refinance to Lower Rate)
- Save $180/month
- Moderate total interest
- Break-even in 24 months
The Growth Plan (Cash-Out Refinance)
- Access $50K for renovations or investments (college savings)
- Payment increases slightly
- Build wealth differently
6. Exploration
Alex clicks each card, uses sliders to adjust assumptions, and compares side-by-side for 10+ minutes.
- Navigator: Tracks heuristic engagement patterns
- Storyboard: Updates visualizations instantly
7. Clarity Moment
Alex understands trade-offs clearly: "Oh, extending the term saves monthly but costs more long-term. Got it."
Advisor: Detects breakthrough moment (confident language). Celebrates understanding
8. Exit
Alex feels empowered: "I know what I need to decide now" (no pressure for Stage 2)
6. Functional Requirements by Tier
6.1 Alpha Scope (Must Have for First Launch)
Goal: Prove the empowerment model works, users complete exploration with confidence.
Alpha will consist only of Stage 1 below. After Stage 1 has gone to market, and we've received user feedback, the development of Stage 2 will commence. Success and go/no go criteria will need to be defined.
Stage 1: Consultation (Anonymous, Free)
Core Features:
- Conversational AI: 3-5 warm questions to discover user intent and goals
- Scenario Generation: Create 3-4 viable refinance paths based on user intent
- Soft Guidance: Highlight which scenario addresses stated priority with reasoning
- Scenario Cards: Visual cards showing monthly payment, total interest, break-even, trade-offs
- Scenarios are presented as parallel paths, not ranked outcomes. The system may explain why a scenario aligns with a stated goal, but does not declare a "best," "optimal," or "recommended" choice.
- Should the user ask explicitly for a recommendation, we must address this carefully.
- Use 3rd party data to enable the expert consultation vs. asking them all the questions in a long form application process. The difference here is that other lenders capture information through long forms; we're trying to capture the least amount of data from the user as possible while still pulling back a bunch of data through third party integrations and insights. Drive conversation vs asking them all this information
- Pictures of the property (from data source). I envision that this makes the recommendations more personal and speeds up trust
- Interactive Calculators: Real-time updates (<200ms) as users adjust assumptions
- Visual Education: 5th-grade level charts explaining amortization, compound interest, trade-offs
- Side-by-Side Comparison: Compare scenarios with transparent trade-offs highlighted
- No Authentication: Completely anonymous, zero friction to explore
- Anonymous refinance scenario generation using self-disclosed data
- Early recommendations generated from minimal inputs, refined dynamically
- Interactive visualizations (sliders, timelines, what-if scenarios)
- Support for all five refinance motivations, including:
- Payment reduction
- Term reduction
- Equity access
- Risk reduction (ARM to Fixed)
- Equity conversion (HELOC and Reverse Mortgage, guidance only)
- These scenarios are educational only, exclude eligibility determination, pricing optimization, or lender readiness, and must clearly state that additional verification is required outside the Clarity Engine.
- Scenario comparison reports (downloadable PDF, unverified data clearly disclosed)
- National-average rate assumptions sourced from non-lender entities (e.g., Freddie Mac)
- Downloadable, shareable PDF reports summarizing explored scenarios, tradeoffs, and assumptions.
- These reports are available once a user opts to save their progress and are explicitly labeled as based on unverified or partially verified data.
Emotional Intelligence (Basic Implementation):
Oracle (Brain):
- Archetype Detection: Classify users as Analytical, Risk-Taking, Emotional, Practical, or Cautious
- Implementation: Rule-based classifier and mortgage intelligence layer using language patterns from MonsterLead insights
- The Oracle does not persist user archetype or preference data across anonymous sessions in Alpha.
- Target: 70%+ confidence by message 5
Navigator (Spine):
- Conversation Orchestration: Manage flow from intent discovery to education to scenario exploration
- Clear Next Steps: Always tell the user "what happens next" or what changes and why based on new information that is adjusted to in real time
Advisor (Heart):
- Emotion Detection: Identify confused, impatient, frustrated, and confident states
- Implementation: Claude API + VADER sentiment analysis for validation
- Tone Adaptation: Adjust warmth/technicality based on emotion + archetype
- Basic Interventions: Simplify on confusion, reassure on anxiety
Storyboard (Skin):
- Visual Clarity: Beautiful, mobile-responsive design
- Interactive Elements: Sliders, toggles, hover states (60fps)
- Chart Library: D3.js or Recharts for visualizations
Explicitly Excluded:
- Lender marketplace or offer matching
- Application or fulfillment workflows
- Voice-first or native app experiences
- Proactive refinance alerts or monitoring
6.2 Post-Alpha Scope
Stage 2: Verification (Authenticated, Free to Consumer)
Core Features:
- Authentication: Email/Google/Apple sign-in with MFA (Supabase Auth)
- Credit Pull: Soft pull via credit bureau API (Experian/Equifax/TransUnion)
- Income/Asset Verification: Truv or Hycron integration, manual fallback
- Cost: PreFi pays for all verification services (part of COGS)
- Live Recalculation: Scenarios update with real data
- Progressive Enhancement: Optional spouse income, 401K balance, additional assets
- Data Privacy: All data encrypted at rest and in transit, NOT shared without explicit permission
- Proactive refinance opportunity alerts
- Expanded contextual education (taxes, insurance, investment overlays)
- Mortgage readiness
- Optional soft credit pull (single bureau) after user opt-in
Stage 3: Marketplace (Real Lender Integration)
Core Features:
- Lender Marketplace & Integration: Prove end-to-end flow with simple API (MonsterLead partner)
- Lender Display: Show lender with rates, fees, estimated closing timeline
- User Selection: User chooses to proceed with lender
- Explicit Consent: Clear authorization flow for data transfer
- Data Package: User's verified data + scenario preference prepared for transfer
- Goal: Validate full funnel works (explore, verify, select)
- Fulfillment orchestration to the lender marketplace.
- Pre-packaged mortgage-ready files.
Platform Summary Deliverables:
- Complete 3-stage platform (Consultation, Verification, Marketplace with ONE lender)
- 4-component emotional intelligence system (basic implementation)
- Mobile-responsive design working on iOS/Android browsers
- Scenario generation engine with 3-4 refinance paths
- Interactive calculators and visualizations
- Credit bureau and verification service integrations
- ONE real lender integration proving end-to-end flow
- Success metrics dashboard and post-session surveys
6.3 Additional Future Enhancement
Advanced Emotional Intelligence (ML-Powered)
Fine-Tuned Archetype Classifier:
- Base Model: Llama 3 8B or GPT-4o-mini
- Training: Fine-tune on labeled MonsterLead conversations
- Validation: 90%+ accuracy on hold-out test set
- Deployment: Fast inference (<100ms) for real-time detection
Enhanced Advisor (Heart):
- Proactive Interventions: Detect crisis moments before user abandons
- Example: 3+ confused signals leads to Simplify explanation leads to Offer visual alternative
- Celebration Moments: Recognize breakthroughs and celebrate understanding
- Example: Confident language after confusion triggers "You got it! That's exactly right."
- Dynamic Tone Adjustment: Adapt warmth/technicality in real-time
ML Feedback Pipeline:
- Track: Which interventions work for which archetypes in which situations
- Learn: User profile + life goal + emotional state = optimal strategy
- A/B Test: Compare intervention strategies, surface improvements
- Deploy: Automated prompt refinement with human review
MonsterLead Data Integration
Data Preparation Phase:
- Process MonsterLead's loan officer conversations
- Clean, normalize, and structure conversation data
- Create taxonomy: mortgage math, borrower math, rapport building, question patterns, emotional signals
- Label conversations by archetype (Analytical, Emotional, Practical, Cautious)
- Filter out "sell" language: identify and remove commission-driven, closing tactics
- Establish boundaries: "advising" vs "selling" distinction
Deliverables:
- Cleaned dataset: 100K+ high-quality conversations
- Taxonomy document: mortgage intelligence categories
- Labeled training set: 20K conversations tagged by archetype
- Boundary guidelines: "sell" vs "advise" training manual
Admin Dashboard & Lender Analytics
Functional Requirements:
- User Analytics: Funnel visualization, completion rates, emotional intelligence metrics
- Conversation Monitoring: Real-time session tracking, intervention analysis
- Lender Performance: Track conversion rates, user satisfaction by lender
- Lead Quality Dashboard: Show lenders what they're getting (emotional journey viz, scenario exploration heatmap, "why they chose you" AI summary)
- Export Capabilities: CSV/Excel reports for financial analysis
- KYC: Name + phone collection
This BECOMES the B2B sales tool: "See how much more you know about this lead?"
Expanded Lender Marketplace:
- Integrate 10+ lenders from the network
- Live Rates: API integrations for real-time rate quotes
- Data Transfer: User authorizes data package transfer to selected lender
- MISMO 3.6 Format: Industry-standard data exchange
Mortgage Motivations and Tools as it Relates to the Borrower Profile:
- More Scenario Types: Cash-out refi, rate-and-term, variable-to-fixed rate, ARM options, changing amortization term (30> or 30<), removing MI, etc.
- Advanced Calculators: Tax implications, investment opportunity cost, break-even analysis
- Richer Visualizations: Amortization schedules, interest vs principal charts, lifetime cost comparisons
Voice Interface
Features:
- Voice Input: Users can ask questions verbally (speech-to-text)
- Voice Output: AI explanations delivered via natural speech (text-to-speech)
- Hybrid Mode: Users can switch between text and voice seamlessly
- Implementation: Web Speech API (browser-native) or dedicated service
- Quality Threshold: 95%+ transcription accuracy, natural-sounding voice
6.4 Full Platform Scope (Scale & White-Label)
Multi-Tenancy & White-Label Capabilities
Functional Requirements:
- Tenant Isolation: Separate data, branding, and configuration per tenant
- White-Label Branding: Custom logos, colors, domain names
- Permission System: Role-based access control (RBAC) for admin users
- Usage-Based Billing: Track and bill based on tenant activity
- Target Customers: Mortgage brokers, financial advisors, credit unions
API Marketplace:
- Public API: Enable third-party integrations
- Developer Portal: Documentation, API keys, usage analytics
- Webhook System: Real-time event notifications
- Rate Limiting: Tiered access based on customer plan
Advanced Features:
- Predictive AI: Anticipate user needs before they ask
- Multi-Product Support: Refinance, purchase, HELOC, commercial
- Advanced Reporting: Custom dashboards, white-label reports for partners
- Enterprise SSO: SAML, OAuth integration for large customers
7. Technical Architecture & Implementation
7.1 Emotional Intelligence: Proposed AI Models
Multi-LLM architecture optimizing for different capabilities:
Primary Conversation Engine:
- Model: Anthropic Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)
- Role: Complex reasoning, conversation orchestration, scenario generation
- Why: Superior at understanding context, adapting tone, maintaining conversational coherence
Emotion Detection (Advisor Heart):
- Primary: Claude API with emotion-detection prompts
- Secondary: VADER Sentiment Analysis (lightweight, fast validation)
- Approach: Dual-model validation (Claude for nuance, VADER for speed)
- Output: 7 emotional states (confused, frustrated, confident, anxious, excited, impatient, overwhelmed, disengaged)
Archetype Detection (Oracle Brain):
- Alpha: Rule-based classifier using language patterns from MonsterLead insights
- Enhancement: Fine-tuned classifier built on MonsterLead 1M+ conversation dataset
- Base Model: Llama 3 8B or GPT-4o-mini (fast, cost-effective inference)
- Training: Supervised learning on labeled conversations (Analytical, Emotional, Practical, Cautious)
- Confidence Threshold: 70%+ by message 5 in Alpha, 90%+ in Enhancement
Financial Calculations:
- Implementation: Python FastAPI backend with mortgage calculation libraries
- Libraries: numpy-financial, custom amortization engine
- Performance: <50ms for scenario calculations
7.2 System Architecture
Frontend:
- Framework: Next.js 14 with React Server Components
- State Management: React Context + Zustand for complex state
- Visualization: D3.js or Recharts for charts, Framer Motion for animations
- Mobile-First: Responsive design, touch-optimized, 60fps interactions
Backend:
- API: FastAPI (Python) for financial calculations and AI orchestration
- Database: Nekt or PostgreSQL with pgvector for semantic search
- Cache: Redis for session management and rate limiting
- Authentication: Supabase Auth (email/social) with MFA
Infrastructure:
- Hosting: Vercel (frontend), AWS/Railway (backend)
- Observability: Datadog (metrics), Sentry (errors), PostHog (analytics)
- CI/CD: GitHub Actions for automated testing and deployment
7.3 MonsterLead Data Integration Strategy
MonsterLead's 1M+ conversations provide the data moat for training emotional intelligence:
What We Learn:
- Mortgage Math: How loan officers explain complex calculations
- Rapport Building: Patterns of trust-building language
- Question Answering: Common borrower questions and effective explanations
- Archetype Patterns: How different personality types ask questions and make decisions
- Incentives: How does the conversation change based on the incentives of the parties
What We Filter Out:
- Sales Language: Remove "close" tactics, urgency pressure, commission-driven recommendations
- Steering Behavior: Identify and eliminate patterns that push specific products
- Create Boundaries: Train on "sell" vs "advising" distinction
Implementation Approach:
- Alpha: Use conversations to inform prompt engineering and tone guidelines
- Enhancement Phase: Formal data preparation, labeling, and fine-tuning
- Full Platform: Continuous learning pipeline improving from every conversation
7.4 Platform Architecture Requirements
CRITICAL: Delivering a PLATFORM, Not Just an App
PreFi is being built as platform infrastructure from day 1. Even though Alpha serves a single tenant (PreFi), the technical architecture supports multi-tenancy, white-label, and API-first design.
Why This Matters: When we scale to Enhancement and Full Platform, we won't need to rebuild. We'll just turn on capabilities that were designed in from the start. This is why the technical team must build the RIGHT foundation now.
Multi-Tenant Database Architecture (Even if Single Tenant in Alpha)
Design for multi-tenancy from day 1:
- Tenant isolation at the database level (tenant_id on all tables)
- Row-level security policies in PostgreSQL
- Separate schema per tenant for data isolation
- Configuration tables supporting tenant-specific settings
Why: Adding tenants later without this foundation requires a database migration nightmare
API-First Design (Even if Only Web UI in Alpha)
Build API before building UI:
- All functionality is accessible via REST/GraphQL API
- UI consumes the same API as external integrations would
- API versioning strategy (v1, v2) from day 1
- Open API documentation auto-generated
Why: Enables mobile apps, partner integrations, and white-label without rebuilding, and our ability to decouple each service/stage for new monetization and partner opportunities.
Configuration System for Branding/White-Label (Even if Not Exposed in Alpha)
Configuration-driven rather than hard-coded:
- Branding assets (logos, colors, fonts) in database, not code
- Feature flags control which capabilities are enabled per tenant
- Email templates, micro-copy, scenario names, all configurable
- Theme system supporting multiple brand identities
Why: White-label customers want their brand, not PreFi's. Build this into the foundation
Horizontal Data Infrastructure Supporting Future Use Cases
Data layer serves multiple products:
- User profiles designed to support refinance AND purchase journeys
- Emotional intelligence data structure works across lending types
- Document processing infrastructure is reusable for any mortgage workflow
- Archetype and emotional state tracking generalizes beyond refinance
Why: Enables future expansion without starting from scratch
Technical Implementation Requirements: To be discussed during discovery
Platform vs Product Thinking: To be discussed during discovery
Success Criteria: When the Enhancement Phase arrives, the technical team can:
- Add new tenant in < 1 day (configuration, not code)
- White-label UI in < 3 days (theme system works)
- Launch API marketplace in < 2 weeks (API already exists)
- Expand to purchase scenarios using same emotional intelligence infrastructure
8. Success Metrics & Validation
8.1 Product Validation Metrics
To be discussed during discovery
- Perceived Neutrality Score: 85%+ of users agree with "I felt informed, not steered toward a specific option."
- A successful session may end without account creation, verification, or lender selection. User clarity and confidence are considered first-order success outcomes.
8.2 Emotional Intelligence Metrics
To be discussed during discovery
8.3 Go/No-Go Decision Criteria
After Beta Testing (50-100 Users):
- If 80%+ of validation targets hit, product validates empowerment model, proceed to Enhancement
- If 60-79%, iterate and re-test with next cohort
- If <60%, fundamental rethink of approach
Key Success Indicators:
- Users complete exploration feeling empowered (>85%)
- No pressure to choose option they're uncomfortable with (>90%)
- Emotional intelligence creates "got me" moments (>75%)
- Conversion to verification shows value (>40%)
9. Privacy, Data Use & Compliance
9.1 Privacy as Core Value
CRITICAL CLARIFICATION:
We sell consumer information when and only when the consumer wants the recommendation fulfilled. We do not sell or share their data/profile if the consumer does not give explicit permission.
Privacy Model:
- Stage 1: Completely anonymous, no data collection
- Stage 2: Data collected for verification, NOT shared with anyone
- Stage 3: User must explicitly authorize data transfer to selected lender
- User Control: Delete data any time through self-service prior to us sharing with fulfillment partners if they give explicit permission to share
This isn't compliance theater. It's competitive advantage.
Security (Build toward industry standards, SOC)
9.2 Freemium Model
Revenue Model:
- Consumer: FREE. No charge for any stage (exploration, verification, marketplace)
- PreFi Pays: Verification costs (credit pull, income verification) are part of COGS
- Lenders Pay: On fulfilled lead (when user authorizes data transfer and selects lender)
- Value Proposition: Higher-quality, better-qualified leads justify premium pricing vs traditional lead gen
Why This Works:
Traditional leads: $50-150, low quality, user called by 5-10 lenders
PreFi leads: $400-800, high quality, high intent, prequalified user selected YOU specifically, complete emotional journey and data structure prepared for optimal deal structure
9.3 Regulatory Compliance
Required Regulations:
- TILA (Truth in Lending Act): Clear disclosure of estimates and assumptions
- FCRA (Fair Credit Reporting Act): Proper handling of credit data, soft pull authorization
- GLBA (Gramm-Leach-Bliley): Financial data privacy and security
- TCPA: No unsolicited contact, explicit opt-in required
Educational Distinction: We educate and empower. We do not provide financial advice.
Required Disclaimer: "Educational information demonstrating financial principles. Not investment or tax advice. Consult licensed professionals."
Prohibited Language:
- "Guaranteed," "risk-free," "you should," "best investment"
Permitted Language:
- "Here's what makes sense," "This strategy demonstrates," "Financial principles suggest"
Appendix A: Why Platform Architecture Matters
Brief Note on Future Capabilities:
The platform architecture we're building for PreFi doesn't just benefit PreFi. It creates the foundation for future lending innovation. The emotional intelligence, scenario generation engine, and data infrastructure become reusable assets.
For PreFi:
- Faster iterations (configuration-driven, not code-driven)
- Better scaling (multi-tenant from day 1)
- White-label revenue (built into architecture)
- API marketplace (API-first design)
For Future:
- Platform enables expansion to new lending products
- Emotional intelligence works across use cases
- Data moat compounds with every conversation
- Infrastructure supports multiple business models
Focus for This PRD: Get PreFi right. Prove the model. Win with refinance. The platform architecture is how we build it correctly from the start. Not premature optimization, but thoughtful foundation.