InsureSense AI - Insurance Domain Sentiment Analyzer

An AI-powered insurance domain sentiment analysis platform that uses a multi-agent system to analyze policyholder communications. Extracts sentiment, emotions, purchase intent, customer persona, journey stage, risk indicators, and policy recommendations. Built with .NET 10 Web API, Angular 21 SPA, and Microsoft Semantic Kernel agent orchestration.

Features

v1 (Legacy - General Purpose)

• Real-time Sentiment Analysis: Analyze any text and get instant results
• Emotion Breakdown: Detailed emotional components (joy, sadness, anger, fear, etc.)
• Confidence Scores: Visual indicators showing confidence levels

v3 (Insurance AI Operations Hub)

• Interactive Landing Page: Public showcase with agent orchestration visualization, provider failover simulation, multimodal pipeline tabs, interactive demo, PII redaction demo
• Multi-Agent AI Analysis: 9-agent pipeline (CTO, BA, Developer, QA, Architect, UX Designer, AI Expert, Claims Triage, Fraud Detection) via Semantic Kernel
• Insurance Context Classification: Claims, policy servicing, billing, agent interaction, underwriting
• Purchase Intent Scoring: 0-100 scale with persona and journey stage detection
• Risk Indicators: Churn risk, complaint escalation risk, fraud indicators
• Claims Triage UI: Submit claims with text + file upload, view severity/urgency/fraud results inline
• Claims History: Filterable/paginated claims table with severity, status, and date range filters
• Fraud Alerts Dashboard: High-risk fraud alert cards with SIU referral indicators
• Provider Health Monitor: Real-time health of 5 LLM providers + 6 multimodal services with auto-refresh
• Chart.js Dashboard: Severity distribution doughnut chart, customer persona bar chart, quick links
• Multimodal Evidence Processing: Upload images (Azure/Cloudflare Vision), audio (Deepgram STT), PDFs (OcrSpace OCR) + NER entity extraction
• PII Redaction: Automatic redaction of SSN, policy numbers, claim numbers, phone, email before external AI calls and DB storage
• Analytics Dashboard: Aggregated metrics, sentiment distribution, persona trends
• Free AI Providers: Groq (primary), Gemini (secondary), Ollama (local fallback) with 5-provider resilient fallback chain
• Persistent Storage: SQLite (development) / Supabase PostgreSQL (production)
• 3 Themes: Dark, semi-dark, and light themes across all 13 components
• WCAG AA Accessibility: axe-core validated on all routes, ARIA attributes, keyboard navigation

v4 (Current — Sprint 4 COMPLETE)

• Document Intelligence RAG: Voyage AI finance-optimized embeddings (1024-dim) + Ollama fallback, insurance section-aware chunking, SQLite vector store with SIMD cosine similarity, upload → OCR → chunk → embed → store, query → embed → vector search → LLM answer with citations
• Customer Experience Copilot: AI-powered SSE streaming chat with insurance CX specialist persona, dual-pass PII redaction (input + output), tone classification, 16-keyword escalation detection, regulatory disclaimer enforcement, CxInteractionRecord audit trail
• Cross-Claim Fraud Correlation: 4-strategy detection (DateProximity, SimilarNarrative, SharedFlags, SameSeverity), claim-type-specific windows (Auto 90d, Property/Liability 180d, WorkersComp 365d), review workflow (Pending/Confirmed/Dismissed)
• V1 PII Fix: Decorator pattern (PiiRedactingSentimentService) wrapping frozen ISentimentService without modifying v1 files
• Per-Endpoint Rate Limiting: analyze 10/min, triage 5/min, fraud 5/min, doc upload 3/min
• MCP Server Integration: Playwright MCP (E2E test generation) + Stitch MCP (design-to-code pipeline)

---

Security & PII Protection

PII protection is not a feature of this platform -- it is the foundational constraint around which every other feature was designed. Insurance communications contain the most sensitive personal data a consumer can share: Social Security numbers, policy identifiers, claim references, contact information. Every architectural decision assumes that raw policyholder text must never reach an external system, a persistent store, or a log file without first passing through the PII redaction pipeline.

Multi-Layer Defense Strategy

PII redaction is enforced at four independent layers, so a failure at any single point does not result in data exposure:

Before / After: PII Redaction in Action

Input (raw policyholder text):

I reported water damage under policy HO-2024-789456. My claim number is
CLM-2024-12345678. My SSN is 987-65-4321. Contact me at jane@email.com
or (555) 123-4567. I'll be contacting the department of insurance.

Output (after PIIRedactionService.Redact()):

I reported water damage under policy [POLICY-REDACTED]. My claim number is
[CLAIM-REDACTED]. My SSN is [SSN-REDACTED]. Contact me at [EMAIL-REDACTED]
or [PHONE-REDACTED]. I'll be contacting the department of insurance.

All sentiment, emotion, and complaint escalation signals are fully preserved -- the LLM receives everything it needs without ever seeing the policyholder's identity.

Redaction Patterns

All patterns use .NET source-generated regex ([GeneratedRegex]) for compile-time optimized matching. Validated by 11 dedicated unit tests.

Local-Only Fallback: Ollama

The fallback chain terminates at Ollama -- a locally hosted LLM that never sends data over the network. For organizations with strict data residency requirements, set AgentSystem:Provider to "Ollama" for zero external data transmission.

Compliance & Data Governance

Audit Trail: Raw text is never logged. Only SHA-256 hashes, timestamps, provider used, and result metadata are recorded.

---

Platform Screens

The Insurance AI Operations Hub ships with 15 routes across claims operations, sentiment analysis, provider monitoring, fraud detection, document intelligence, and customer experience. Built with a glassmorphism design system featuring translucent cards with backdrop-filter: blur(12px), indigo-to-purple gradient accents, and staggered entrance animations.

Design System: 3-theme toggle (Dark / Semi-Dark / Light), glassmorphism cards with 12px blur, responsive Tailwind breakpoints, WCAG AA accessibility, prefers-reduced-motion support.

Landing Page (/)

7-section scrolling showcase with animated gradient orbs, headline with gradient text, and CTA buttons. Agent orchestration animation, provider failover simulation, multimodal pipeline tabs, interactive demo, PII redaction toggle, stats grid, and tech badges.

Insurance Analyzer (/insurance)

Multi-agent insurance analysis across 7 dimensions (sentiment, emotions, purchase intent, persona, journey stage, risk indicators, recommendations). Quick templates, 10K-char input, session-cached history. Results display with confidence bars and color-coded risk badges.

Analytics Dashboard (/dashboard)

KPI metric cards (Total Analyses, Avg Purchase Intent, Avg Sentiment, High Risk Alerts) + claims KPIs. Chart.js doughnut (sentiment distribution) and horizontal bar chart (customer personas). Quick-action navigation cards to Claims Triage, History, Provider Health, and Fraud Alerts.

Claims Triage (/claims/triage)

Glassmorphism form with resizable text area (10K char limit), interaction type dropdown, 4 quick template buttons, and drag-and-drop evidence upload zone (images/audio/PDFs). Submit triggers a 5-phase loading animation with gradient progress bar.

Inline results: severity badge (color-coded Critical/High/Medium/Low), fraud score gauge (gradient green-to-red), recommended actions accordion with priority badges, fraud flag chips.

Claims History (/claims/history)

Filterable/paginated table with severity, status, and date range filters. Color-coded badges per row for severity (Critical=rose, High=orange, Medium=amber, Low=emerald), urgency, fraud score, and status. Click any row to view full claim details.

Claim Detail (/claims/:id)

Deep view: triage summary grid, fraud gauge, expandable recommended actions, evidence viewer with multimodal results (vision/STT/OCR), and on-demand fraud analysis button.

Provider Health (/dashboard/providers)

Horizontal fallback chain visualization (Groq→Cerebras→Mistral→Gemini→OpenRouter→OpenAI→Ollama) color-coded by status (emerald=Healthy, amber=Degraded, rose=Down). LLM provider cards with availability, consecutive failures, cooldown countdown. Multimodal services grid (Deepgram STT, Azure Vision, Cloudflare Vision, OCR.space, HuggingFace NER). Auto-refreshes every 30 seconds.

Fraud Alerts (/dashboard/fraud)

Summary stats: Critical Risk, High Risk, Avg Score, SIU Referrals. Alert cards sorted by fraud score with left-border accent coloring, SIU Referral flags, fraud gauge bars, claim info grid, category-colored fraud flag badges, and "View Claim" / "Deep Analysis" action buttons.

Document Upload (/documents/upload)

Drag-and-drop file upload for RAG pipeline processing. Category selector (Policy/Claim/Endorsement/Correspondence/Other), 5MB limit, PDF/PNG/JPEG/TIFF. Multi-phase loading animation (OCR → Chunking → Embedding). Result card with document ID, page count, chunk count, embedding provider.

Document Query (/documents/query)

RAG-powered natural language Q&A across uploaded insurance documents. Optional document scope filter, 2000-char question input. Answer display with confidence gauge (green/yellow/red), expandable citations with similarity scores, LLM provider + elapsed time metadata.

Document Detail (/documents/:id)

Full document metadata viewer: file name, MIME type, category badge, status, page/chunk counts. Scrollable chunk browser with section names and token counts. Inline Q&A scoped to document. Delete with confirmation modal.

CX Copilot (/cx/copilot)

AI-powered customer experience chat with SSE streaming. User/assistant message bubbles, incremental token display with blinking cursor. Tone badges (Professional/Empathetic/Urgent/Informational), escalation warnings, regulatory disclaimers. Optional claim context input.

Fraud Correlations (/fraud/correlations/:claimId)

Cross-claim fraud pattern analysis. Summary stats (total correlations, avg score, pending/confirmed counts). Correlation cards with split-card design (source vs correlated claim), strategy badges (DateProximity/SimilarNarrative/SharedFlags/SameSeverity), score gauge. Status filter tabs, review workflow (Confirm/Dismiss with reason).

Sentiment Analyzer v1 (/sentiment)

Legacy general-purpose sentiment analysis with text input, confidence bar, and emotion breakdown chart.

---

Tech Stack

Backend

• .NET 10 Web API (C# 13, net10.0)
• Microsoft Semantic Kernel 1.71.0 (Agent orchestration)
• MediatR 14.0 (CQRS pattern)
• Entity Framework Core 10 (SQLite / PostgreSQL)
• ASP.NET Core Minimal API + Controllers hybrid

Frontend

• Angular 21.1.0 (standalone components, signals)
• TypeScript 5.9.2 (strict mode)
• Tailwind CSS 3.4.17
• Vitest 4.0.8 (testing)

AI Providers (Priority Order — 5-Provider Fallback Chain)

1. Groq - Llama 3.3 70B, fastest inference (250 req/day free)
2. Mistral - Mistral Small, secondary (500K tokens/month free)
3. Gemini - gemini-2.5-flash, best quality (60 req/min free)
4. OpenRouter - Multi-model router ($1 free credit)
5. Ollama - llama3.2, local inference (unlimited, PII-safe)

Multimodal Services

Testing

• Backend: xUnit 2.9.3 + Moq 4.20.72 (461 tests)
• Frontend unit: Vitest 4.0.8 via Angular CLI (235 tests across 28 spec files)
• E2E: Playwright 1.58+ with @axe-core/playwright (357 tests across 16 spec files)
• Total: 1,053 tests, 0 failures

Prerequisites

• .NET 10 SDK
• Node.js 22+ and npm 11+
• At least one AI provider API key:
  • Groq API Key (free at console.groq.com) -- recommended
  • Gemini API Key (free at aistudio.google.com)
  • Ollama installed locally (free at ollama.com)
  • OpenAI API Key (for v1 legacy only)

Setup Instructions

1. Clone the Repository

cd SentimentAnalyzer

2. Backend Setup

cd Backend

Configure API Keys

Create or edit appsettings.Development.json with your provider keys:

{
  "OpenAI": {
    "ApiKey": "your-openai-api-key-here",
    "Model": "gpt-4o-mini"
  },
  "AgentSystem": {
    "Provider": "Groq",
    "Groq": {
      "ApiKey": "your-groq-api-key-here",
      "Model": "llama-3.3-70b-versatile",
      "Endpoint": "https://api.groq.com/openai/v1"
    },
    "Gemini": {
      "ApiKey": "your-gemini-api-key-here",
      "Model": "gemini-2.5-flash",
      "Endpoint": "https://generativelanguage.googleapis.com/v1beta/openai/"
    },
    "Ollama": {
      "Model": "llama3.2",
      "Endpoint": "http://localhost:11434/v1"
    }
  },
  "Database": {
    "Provider": "Sqlite"
  },
  "ConnectionStrings": {
    "DefaultConnection": "Data Source=insurance_analysis.db"
  }
}

Important: Never commit appsettings.Development.json to version control (it is gitignored).

Install Dependencies and Run

dotnet restore
dotnet run

The API will start at http://localhost:5143

3. Frontend Setup

Open a new terminal:

cd Frontend/sentiment-analyzer-ui
npm install
npm start

The app will open at http://localhost:4200

4. Run Tests

# Backend tests (461 tests)
cd SentimentAnalyzer/Tests
dotnet test

# Frontend unit tests (235 tests - must use Angular CLI, not direct vitest)
cd SentimentAnalyzer/Frontend/sentiment-analyzer-ui
npx ng test --watch=false

# E2E tests (357 tests - requires frontend dev server running)
cd SentimentAnalyzer/Frontend/sentiment-analyzer-ui
npm run e2e

Usage

1. Navigate to http://localhost:4200 in your browser
2. Use the navigation bar to switch between:
   • Home (/) - Interactive landing page showcasing the platform
   • Sentiment Analyzer (/sentiment) - v1 general sentiment analyzer
   • Insurance Analyzer (/insurance) - v2 multi-agent insurance analysis
   • Claims Triage (/claims/triage) - Submit claims for AI triage
   • Claims History (/claims/history) - Browse and filter past claims
   • Dashboard (/dashboard) - Analytics, charts, and quick links
   • Provider Health (/dashboard/providers) - Real-time AI provider health
   • Fraud Alerts (/dashboard/fraud) - High-risk fraud alert monitoring
   • Document Upload (/documents/upload) - Upload documents for RAG processing
   • Document Query (/documents/query) - Ask questions about uploaded documents
   • CX Copilot (/cx/copilot) - AI-powered customer experience chat
   • Login (/login) - Supabase authentication (optional)
3. On the Insurance Analyzer page:
   • Enter policyholder text or use a sample template
   • Select the interaction type (General, Email, Call, Chat, Review, Complaint)
   • Click Analyze to run the multi-agent pipeline
   • View sentiment, emotions, purchase intent, persona, risk indicators, and recommendations
4. On the Claims Triage page:
   • Enter claim description text and optionally upload evidence files (images, audio, PDFs)
   • Select interaction type and click Triage Claim
   • View severity, urgency, fraud score, recommended actions, and fraud flags inline

API Endpoints

v1 (Legacy - Frozen)

v1 Request:

{
  "text": "Your text here"
}

v1 Response:

{
  "sentiment": "Positive",
  "confidenceScore": 0.95,
  "explanation": "The text expresses strong positive emotions...",
  "emotionBreakdown": {
    "joy": 0.8,
    "excitement": 0.6,
    "satisfaction": 0.7
  }
}

v2 (Insurance Domain)

Claims & Fraud Pipeline (Sprint 2)

Document Intelligence RAG (Sprint 4 Week 2)

CX Copilot + Fraud Correlation (Sprint 4 Week 3)

v2 Request:

{
  "text": "I reported water damage on Jan 15. It has been 3 weeks with no response. Policy HO-2024-789456.",
  "interactionType": "Complaint"
}

v2 Response:

{
  "sentiment": "Negative",
  "confidenceScore": 0.92,
  "explanation": "Customer is expressing frustration with claim processing delays...",
  "emotionBreakdown": { "frustration": 0.85, "anger": 0.70 },
  "insuranceAnalysis": {
    "purchaseIntentScore": 15,
    "customerPersona": "ClaimFrustrated",
    "journeyStage": "ActiveClaim",
    "riskIndicators": {
      "churnRisk": "High",
      "complaintEscalationRisk": "High",
      "fraudIndicators": "None"
    },
    "policyRecommendations": [
      { "product": "Claims Fast-Track", "reasoning": "Expedited claims processing to reduce churn" }
    ],
    "interactionType": "Complaint",
    "keyTopics": ["claim delay", "water damage", "no response"]
  },
  "quality": {
    "isValid": true,
    "qualityScore": 92,
    "issues": [],
    "suggestions": [],
    "warnings": []
  }
}

Project Structure

SentimentAnalyzer/
├── Backend/
│   ├── Controllers/SentimentController.cs     # v1 API (FROZEN - never modify)
│   ├── Endpoints/
│   │   ├── InsuranceEndpoints.cs              # v2 Minimal API + MediatR
│   │   ├── ClaimsEndpoints.cs                 # Claims triage + evidence upload endpoints
│   │   ├── FraudEndpoints.cs                  # Fraud analysis + alerts endpoints
│   │   └── ProviderHealthEndpoints.cs         # Provider health monitoring endpoint
│   ├── Features/
│   │   ├── Insurance/
│   │   │   ├── Commands/AnalyzeInsuranceCommand.cs
│   │   │   └── Queries/ (GetDashboardQuery, GetHistoryQuery)
│   │   ├── Claims/
│   │   │   ├── Commands/ (TriageClaimCommand, UploadClaimEvidenceCommand)
│   │   │   └── Queries/ (GetClaimQuery, GetClaimsHistoryQuery)
│   │   ├── Fraud/
│   │   │   ├── Commands/AnalyzeFraudCommand.cs
│   │   │   └── Queries/ (GetFraudScoreQuery, GetFraudAlertsQuery)
│   │   └── Health/Queries/GetProviderHealthQuery.cs
│   ├── Data/
│   │   ├── InsuranceAnalysisDbContext.cs       # EF Core DbContext (6 DbSets)
│   │   ├── IAnalysisRepository.cs             # Sentiment analysis repository
│   │   ├── SqliteAnalysisRepository.cs        # SQLite implementation
│   │   ├── IClaimsRepository.cs               # Claims domain repository
│   │   ├── SqliteClaimsRepository.cs          # Claims SQLite implementation
│   │   └── Entities/
│   │       ├── AnalysisRecord.cs              # Sentiment analysis entity
│   │       ├── ClaimRecord.cs                 # Claims triage entity
│   │       ├── ClaimEvidenceRecord.cs         # Multimodal evidence entity
│   │       └── ClaimActionRecord.cs           # Recommended actions entity
│   ├── Models/                                # Request/Response DTOs
│   │   ├── SentimentRequest/Response.cs       # v1 (frozen)
│   │   ├── InsuranceAnalysisResponse.cs       # v2
│   │   ├── ClaimTriageRequest/Response.cs     # Claims triage
│   │   ├── FraudAnalysisResponse.cs           # Fraud scoring
│   │   ├── ProviderHealthResponse.cs          # Provider health
│   │   └── PaginatedResponse.cs               # Generic pagination wrapper
│   ├── Services/
│   │   ├── PIIRedactionService.cs             # PII redaction (source-generated regex)
│   │   ├── Claims/
│   │   │   ├── ClaimsOrchestrationService.cs  # Claims triage facade
│   │   │   └── MultimodalEvidenceProcessor.cs # MIME routing (vision/STT/OCR + NER)
│   │   ├── Fraud/FraudAnalysisService.cs      # Fraud scoring facade
│   │   ├── ISentimentService.cs               # v1 (frozen)
│   │   └── OpenAISentimentService.cs          # v1 (frozen)
│   ├── Middleware/GlobalExceptionHandler.cs    # IExceptionHandler
│   └── Program.cs                             # DI, middleware, endpoint registration
│
├── Agents/
│   ├── Configuration/                         # AgentSystemSettings, AgentConfiguration
│   ├── Definitions/
│   │   ├── AgentDefinitions.cs                # System prompts (9 agents)
│   │   └── AgentRole.cs                       # Agent role enum (9 roles)
│   ├── Orchestration/
│   │   ├── InsuranceAnalysisOrchestrator.cs   # Profile-aware AgentGroupChat pipeline
│   │   ├── AgentSelectionStrategy.cs          # Deterministic turn-taking
│   │   └── AnalysisTerminationStrategy.cs     # ANALYSIS_COMPLETE detection
│   ├── Plugins/                               # Semantic Kernel plugins
│   └── Models/
│       ├── AgentAnalysisResult.cs             # Agent output (incl. ClaimTriage + FraudAnalysis)
│       ├── ClaimTriageDetail.cs               # Claims triage agent output model
│       └── FraudAnalysisDetail.cs             # Fraud detection agent output model
│
├── Domain/
│   ├── Enums/                                 # SentimentType, CustomerPersona, InteractionType, etc.
│   └── Models/                                # Shared domain models
│
├── Frontend/sentiment-analyzer-ui/
│   └── src/app/
│       ├── components/ (18 total)
│       │   ├── landing/                       # Public landing page (interactive showcase)
│       │   ├── sentiment-analyzer/            # v1 general analyzer (legacy)
│       │   ├── insurance-analyzer/            # v2 insurance analysis UI (signals, timer, phases)
│       │   ├── dashboard/                     # Analytics dashboard (Chart.js charts, quick links)
│       │   ├── claims-triage/                 # Claims triage form + result display
│       │   ├── claim-result/                  # Claim detail view by ID
│       │   ├── evidence-viewer/               # Multimodal evidence child component
│       │   ├── claims-history/                # Filterable/paginated claims table
│       │   ├── provider-health/               # LLM + multimodal service health monitor
│       │   ├── fraud-alerts/                  # High-risk fraud alert cards
│       │   ├── document-upload/               # Document RAG upload (drag-drop, OCR progress)
│       │   ├── document-query/                # RAG document Q&A with citations
│       │   ├── document-result/               # Document detail + chunk browser
│       │   ├── cx-copilot/                    # CX Copilot SSE streaming chat
│       │   ├── fraud-correlation/             # Cross-claim fraud pattern analysis
│       │   ├── login/                         # Supabase auth login
│       │   └── nav/                           # Navigation bar (theme toggle, mobile menu)
│       ├── services/
│       │   ├── sentiment.service.ts           # v1 HTTP client
│       │   ├── insurance.service.ts           # v2 API client (inject() pattern)
│       │   ├── claims.service.ts              # Claims/fraud/health API client (8 methods)
│       │   ├── auth.service.ts                # Supabase auth (signals)
│       │   └── theme.service.ts               # Theme switching (dark/semi-dark/light)
│       ├── models/
│       │   ├── sentiment.model.ts             # v1 interfaces
│       │   ├── insurance.model.ts             # v2 interfaces (QualityDetail, QualityIssue, 14+ types)
│       │   └── claims.model.ts                # Claims/fraud/health interfaces
│       ├── guards/
│       │   ├── auth.guard.ts                  # Route protection (CanActivateFn)
│       │   └── guest.guard.ts                 # Guest-only routes
│       └── interceptors/
│           ├── auth.interceptor.ts            # JWT header injection
│           └── error.interceptor.ts           # 401/403 redirect handling
│   └── e2e/ (12 spec files, 263 tests)
│       ├── fixtures/mock-data.ts              # Realistic insurance mock API responses
│       ├── helpers/api-mocks.ts               # page.route() interceptors
│       ├── claims-triage.spec.ts              # Claims triage E2E tests
│       ├── claims-detail.spec.ts              # Claim detail E2E tests
│       ├── claims-history.spec.ts             # Claims history E2E tests
│       ├── provider-health.spec.ts            # Provider health E2E tests
│       ├── fraud-alerts.spec.ts               # Fraud alerts E2E tests
│       └── ... (7 more spec files)
│
├── Tests/ (461 tests)
│   ├── SentimentControllerTests.cs            # v1 regression (9 tests - FROZEN)
│   ├── InsuranceAnalysisControllerTests.cs    # CQRS handler tests (27 tests incl. 7 MapQuality)
│   ├── PIIRedactionTests.cs                   # PII redaction tests (11 tests)
│   ├── UnitTest1.cs                           # Placeholder (1 test)
│   ├── OrchestrationProfileFactoryTests.cs    # Profile → agent mapping tests
│   ├── ProviderConfigurationTests.cs          # LLM provider config tests
│   ├── ResilientKernelProviderTests.cs        # 5-provider fallback chain tests
│   ├── HuggingFaceNerServiceTests.cs          # NER entity extraction tests
│   ├── DeepgramServiceTests.cs                # Speech-to-text tests
│   ├── AzureVisionServiceTests.cs             # Azure Vision image analysis tests
│   ├── CloudflareVisionServiceTests.cs        # Cloudflare Vision fallback tests
│   ├── OcrSpaceServiceTests.cs                # OCR document extraction tests
│   ├── CriticalFixTests.cs                    # Sprint 1 critical fix tests
│   ├── FinBertSentimentServiceTests.cs        # FinBERT pre-screening tests (8 tests)
│   ├── AnalyzeInsurancePreScreenTests.cs      # FinBERT handler integration (6 tests)
│   ├── ClaimsOrchestrationServiceTests.cs     # Claims triage facade tests (10 tests)
│   ├── MultimodalEvidenceProcessorTests.cs    # MIME routing + vision fallback tests (10 tests)
│   ├── FraudAnalysisServiceTests.cs           # Fraud scoring + SIU referral tests
│   ├── TriageClaimHandlerTests.cs             # Claims command handler tests
│   ├── UploadClaimEvidenceHandlerTests.cs     # Evidence upload handler tests
│   ├── ClaimsRepositoryTests.cs               # Claims DB persistence tests
│   ├── GetClaimHandlerTests.cs                # Claims query handler tests (4 tests)
│   ├── FraudCommandsTests.cs                  # Fraud command/query tests
│   └── ProviderHealthTests.cs                 # Provider health endpoint tests
│
├── PROJECT_CONTEXT.md
├── SPRINT-ROADMAP.md
├── REVIEW.md
├── QA_REPORT.md
└── README.md (this file)

Configuration

AI Provider Selection

Set AgentSystem:Provider in appsettings.json to switch providers:

• "Groq" (default) - Fastest, recommended for development
• "Gemini" - Higher quality analysis
• "Ollama" - Local inference, no API key needed, PII-safe
• "OpenAI" - Legacy, uses existing credits

Database Provider

Set Database:Provider in appsettings.json:

• "Sqlite" (default) - Local file-based, zero setup
• "PostgreSQL" - Supabase cloud (500MB free tier)

Supabase (Production)

{
  "Database": { "Provider": "PostgreSQL" },
  "ConnectionStrings": {
    "DefaultConnection": "Host=db.your-project.supabase.co;Port=5432;Database=postgres;Username=postgres;Password=your-password"
  },
  "Supabase": {
    "Url": "https://your-project.supabase.co",
    "JwtSecret": "your-jwt-secret"
  }
}

Frontend API URL

The v2 Insurance service uses environment configuration. Update Frontend/sentiment-analyzer-ui/src/environments/environment.ts:

export const environment = {
  production: false,
  apiUrl: 'http://localhost:5143'
};

Troubleshooting

CORS Issues

CORS is configured in Program.cs to allow http://localhost:4200. If using a different frontend port, update the CORS policy in Backend/Program.cs.

AI Provider Errors

• Groq 429: Free tier limit reached (250 req/day). Switch to Gemini or Ollama in appsettings.json.
• Gemini 429: Rate limit (60 req/min). Add delay between requests or switch provider.
• Ollama connection refused: Ensure Ollama is running (ollama serve) on port 11434.
• OpenAI errors: Verify API key and credit balance.

Frontend Test Runner

Tests must be run via Angular CLI, not directly with Vitest:

# Correct
npx ng test --watch=false

# Incorrect - will fail with "describe is not defined"
npx vitest run

Port Already in Use

• Backend: Change port in Backend/Properties/launchSettings.json
• Frontend: Run with custom port: ng serve --port 4300
• Ollama: Default port 11434

Backend Build While Running

Backend DLLs are locked while the API server is running. Stop the server before rebuilding.

Ports

License

This project is licensed under the MIT License.

Contributing

Contributions welcome! Please open an issue or submit a pull request.

Contact

For questions or feedback, please open an issue in the repository.