Lead Intelligence for Wealthsimple Advisors

An AI-powered lead qualification and scoring system that transforms client transaction data into prioritized advisor opportunities. Built for Wealthsimple's financial advisory team.

Live app: wealthsimple-lead-gen.onrender.com

The advisor's time goes to what matters -- the conversation.

The Problem

Wealthsimple now holds clients' bank accounts. Transaction data is a goldmine of cross-sell signals -- $2,000/month PADs flowing to TD Direct Investing, mortgage renewals at Scotiabank, idle cash sitting uninvested -- but advisors managing 200+ accounts have no systematic way to find them. In practice, these signals go unnoticed.

The Solution

This system scans client banking activity, identifies high-value opportunities using a hybrid AI approach (deterministic rules + LLM synthesis), scores leads on a 5-dimension qualification framework, and surfaces them as ranked, actionable recommendations -- each with a plain-English explanation of why this client was flagged.

Hours of manual forensic work become seconds of informed review.

---

Key Features

Dashboard

Total leads, high-priority count, average composite score, opportunity breakdown (investing vs lending), tier distribution, and top signal types -- all at a glance.

Lead Qualification Scoring (5-Dimension System)

Every lead is scored 0-100 across five dimensions with vertical-specific weights:

Dimension	What It Measures
Demographic Fit	Age, income, employment stability, location, credit score
Financial Qualification	DTI ratio, savings patterns, defaults, collateral potential
Behavioral Engagement	Web/app activity, email engagement, form submissions, branch visits
Intent Signals	Applications started, rate comparisons, life events, competitor mentions
Compliance Readiness	Identity verified, sanctions/PEP screening, KYC/AML status

Tier Thresholds:

• Tier A (80-100): Sales-Ready -- route to senior advisor within 4 hours
• Tier B (60-79): Sales-Qualified -- 24-48 hour follow-up
• Tier C (40-59): Marketing-Qualified -- automated nurture
• Tier D (0-39): Unqualified -- archive

AI Analysis Layer

• Deterministic Rules: Transaction pattern matching for competitor flows, mortgage payments, insurance premiums, income changes, idle cash, and life events
• LLM Synthesis: Claude's tool-use feature produces structured analysis -- score, signals, summary, reasoning, and recommended actions
• Human-in-the-Loop: AI surfaces recommendations; advisors make final contact and product decisions. No lead progresses without human judgment

Meeting Notes Analysis

Advisors upload notes from client meetings (drag-drop .txt, .docx, .xlsx, .pdf). The AI extracts new signals, generates insights that supplement transaction-based analysis, and adjusts lead scores accordingly.

Competitor Intelligence

One-click competitive analysis comparing Wealthsimple vs. detected competitor institutions, with talking points tailored to the specific client's products and dollar amounts.

AI Resilience Architecture

The system never fails due to LLM availability. A multi-provider fallback chain ensures uptime:

Claude Sonnet 4 (primary, best quality)
  --> retry once on transient errors (429/500/529)
Claude Haiku 4.5 (fallback, fast)
  --> retry once on transient errors
Groq Llama 3.3 70B (last resort, open-source)
  --> Zod schema validation
  --> if invalid: Claude auto-repairs the response

All responses are validated against Zod schemas before reaching the database.

---

Tech Stack

Layer	Technology
Framework	Next.js 16.1.6 (App Router)
Frontend	React 19, Tailwind CSS 4, TypeScript 5
Database	SQLite via better-sqlite3
Primary AI	Anthropic Claude (Sonnet 4 + Haiku 4.5)
Fallback AI	Groq (Llama 3.3 70B via OpenAI SDK)
Validation	Zod 4.3.6 (runtime schema validation)
File Parsing	pdf-parse, SheetJS (xlsx), UTF-8 passthrough

---

Getting Started

Prerequisites

• Node.js 20+
• An Anthropic API key (required)
• A Groq API key (optional -- fallback only)

Installation

git clone https://github.com/stsbrega/Lead-Intelligence-Platform.git
cd wealthsimple-lead-gen
npm install

Environment Setup

Copy the example environment file and add your API keys:

cp .env.local.example .env.local

Edit .env.local:

ANTHROPIC_API_KEY=sk-ant-your-key-here
GROQ_API_KEY=gsk_your-key-here        # optional

Run Development Server

npm run dev

Open http://localhost:3000 for local development, or visit the live deployment.

Build for Production

npm run build
npm run start

---

Project Structure

src/
├── app/                        # Next.js App Router
│   ├── api/                    # REST API routes
│   │   ├── analyze/            # Trigger AI analysis for a client
│   │   ├── analyze-notes/      # Analyze advisor meeting notes
│   │   ├── create-lead-from-notes/  # New lead from unstructured text
│   │   ├── competitor-compare/ # Competitive intelligence
│   │   ├── dashboard/          # Dashboard metrics
│   │   ├── lead-counts/        # Counts by status/tier
│   │   └── leads/              # List + detail + status updates
│   ├── leads/
│   │   ├── page.tsx            # Lead list with filtering
│   │   ├── new/page.tsx        # "Add from Notes" upload page
│   │   └── [id]/page.tsx       # Lead detail (scoring, analysis, decisions)
│   └── methodology/            # Scoring methodology documentation
├── components/
│   ├── layout/                 # Sidebar, Providers
│   ├── leads/                  # AdvisorDecisionPanel, DropZones, Signals
│   └── ui/                     # Card, Badge, ScoreBar, MetricCard
├── lib/
│   ├── ai/                     # AI layer
│   │   ├── client.ts           # Resilience chain (Sonnet -> Haiku -> Groq -> self-heal)
│   │   ├── schemas.ts          # Zod validation schemas
│   │   ├── engine.ts           # Transaction analysis orchestration
│   │   ├── lead-from-notes.ts  # Notes -> structured lead
│   │   ├── notes-analyzer.ts   # Notes -> supplemental insights
│   │   ├── groq-client.ts      # Groq/OpenAI-compatible wrapper
│   │   ├── self-heal.ts        # Auto-repair malformed responses
│   │   ├── signals.ts          # Deterministic signal detection
│   │   └── prompts.ts          # System & user prompt templates
│   ├── data/
│   │   ├── db.ts               # SQLite connection
│   │   └── schema.sql          # Database schema
│   ├── scoring/                # 5-dimension qualification scoring
│   │   ├── config.ts           # Weights, thresholds, criteria
│   │   ├── scorer.ts           # Core dimension scoring logic
│   │   └── compute.ts          # Server-side score computation
│   └── file-parser.ts          # DOCX/XLSX/PDF/TXT extraction
└── types/
    └── index.ts                # Central TypeScript interfaces

---

API Routes

Method	Endpoint	Purpose
POST	/api/analyze	Trigger AI analysis for a client
POST	/api/analyze-notes	Analyze advisor meeting notes
POST	/api/create-lead-from-notes	Generate new lead from unstructured text
POST	/api/competitor-compare	Competitive intelligence comparison
GET	/api/dashboard	Dashboard metrics
GET	/api/lead-counts	Lead counts by status and tier
GET	/api/leads	List all leads
GET	/api/leads/[id]	Single lead detail + transactions + analysis
PATCH	/api/leads/[id]	Update lead status and advisor notes

---

Signal Detection

The system detects these signal types from transaction data:

Signal	Description	Example
competitor_rrsp	Monthly contributions to competitor RRSP	$500/mo PAD to TD Direct Investing
competitor_tfsa	TFSA flows to competitor platforms	$300/mo to RBC Direct Investing
competitor_insurance	Insurance premiums at competitors	$200/mo to Sun Life
mortgage_refinance	Mortgage payments to another lender	$2,100/mo to Scotiabank Mortgage
income_change	Salary increase detected (>10%)	Direct deposit jumped from $4k to $5.2k
large_balance_idle	>$25K sitting uninvested	$42,000 in chequing, no investment activity
life_event	Large one-time deposit (>$10K)	$85,000 inheritance deposit

---

Design Decisions

1. Human-in-the-Loop: AI surfaces opportunities; advisors own all contact and product decisions (CIRO KYC/suitability compliance)
2. Deterministic + LLM Hybrid: Rules identify patterns predictably; Claude synthesizes context and nuance
3. Compliance Gate: Separate compliance sub-score auto-disqualifies sanctioned entities
4. Signal Decay: Behavioral signals lose weight over time (configurable half-life)
5. Vertical-Specific Weights: Scoring adapts to product category (retail banking, wealth management, mortgage, commercial)
6. Multi-Provider Resilience: Three LLM providers with automatic fallback and self-healing validation
7. Immutable Audit Trail: All advisor decisions logged with timestamps

---

Deployment

The app is deployed on Render. Required environment variables:

Variable	Required	Description
ANTHROPIC_API_KEY	Yes	Claude API key for primary AI analysis
GROQ_API_KEY	No	Groq API key for last-resort fallback

---

Demo Materials

• demo/ -- Sample data files (CSV bank statements, meeting notes)
• docs/demo-script.md -- Silent video demo recording guide
• docs/written-explanation.md -- Executive summary of AI responsibility boundaries