Skip to content

priyankshusheet/ClauseWiseAI

BranchesTags

Repository files navigation

ClauseWise AI — VI-Sem Minor Project Report

AI-Powered Financial and Legal Document Intelligence Platform

Abstract

The proliferation of complex financial documents—loan agreements, insurance policies, credit card terms—has created a significant comprehension gap for consumers and professionals alike. This project presents ClauseWise AI, a web-based intelligent document analysis platform that leverages Artificial Intelligence, Optical Character Recognition (OCR), and Natural Language Processing (NLP) to simplify, analyze, and assess risk in financial and legal documents. Built on a modern technology stack comprising React, TypeScript, Tailwind CSS, and Supabase, the system provides features including multi-format document upload with PDF.js-based text extraction, Tesseract.js-powered multi-language OCR, AI-driven risk scoring with clause-level breakdown, real-time AI chat with document context, side-by-side document comparison, portfolio-level risk aggregation, and a 30-day financial literacy course. A multi-provider AI inference strategy ensures high availability using a fallback chain across Gemini, OpenAI GPT-4o-mini, Groq Llama-3.3-70b, and Cohere Command-R-Plus. The platform is deployed as a Progressive Web Application (PWA) with offline capability, service workers, and OTP-based authentication. Results demonstrate that ClauseWise AI significantly reduces the time required to comprehend financial documents while providing actionable, explainable risk intelligence grounded in identifiable document text.

Keywords: Document Intelligence, Risk Analysis, NLP, OCR, AI Chat, Financial Literacy, PWA, React, Supabase

Acknowledgement

We express our sincere gratitude to our project guide and the faculty of the Department of Computer Science and Engineering for their invaluable guidance and encouragement throughout the development of this project.

We extend our appreciation to our institution for providing the necessary infrastructure and resources. We are also thankful to the open-source communities behind React, Supabase, Tesseract.js, PDF.js, and the various AI model providers whose tools and platforms made this project possible.

Finally, we acknowledge the contributions of all team members and peers who provided feedback and testing support during the development lifecycle.

Table of Contents

  1. Abstract
  2. Acknowledgement
  3. Acronyms
  4. Nomenclature
  5. List of Figures
  6. List of Tables
  7. Chapter 1: Introduction
  8. Chapter 2: Literature Survey
  9. Chapter 3: Problem Formulation
  10. Chapter 4: Requirement Analysis
  11. Chapter 5: System Design
  12. Chapter 6: Proposed Methodology
  13. Chapter 7: Results & Discussion
  14. Chapter 8: Conclusion and Future Work
  15. References

Acronyms

Acronym Full Form
AI Artificial Intelligence
API Application Programming Interface
CORS Cross-Origin Resource Sharing
CSS Cascading Style Sheets
DFD Data Flow Diagram
DOM Document Object Model
GDPR General Data Protection Regulation
HSL Hue, Saturation, Lightness
HTML HyperText Markup Language
HTTP HyperText Transfer Protocol
JSON JavaScript Object Notation
JWT JSON Web Token
LLM Large Language Model
NLP Natural Language Processing
OCR Optical Character Recognition
OG Open Graph
OTP One-Time Password
PDF Portable Document Format
PWA Progressive Web Application
REST Representational State Transfer
RLS Row-Level Security
SDK Software Development Kit
SEO Search Engine Optimization
SPA Single-Page Application
SQL Structured Query Language
TSX TypeScript XML
UI User Interface
UML Unified Modeling Language
URL Uniform Resource Locator
UUID Universally Unique Identifier
UX User Experience

Nomenclature

Symbol / Term Description
Risk Score (R) A numerical value (0–100) representing the aggregate risk level of a document, computed from clause-level analysis
Confidence Threshold (θ) Minimum OCR confidence value (0–1) below which extracted text is flagged for manual review
Processing Time (T_p) Time in milliseconds from document upload to completion of analysis
Session Duration (S_d) Duration in seconds of a user's active session
Fallback Chain Ordered sequence of AI providers attempted when the primary provider fails
Token A unit of text processed by an LLM; also refers to authentication tokens (JWT)
Edge Function Serverless function deployed on Supabase infrastructure
Design Token Semantic CSS variable used to maintain visual consistency across themes

List of Figures

Figure No. Caption
Fig. 5.1 Overall System Architecture Diagram
Fig. 5.2 Data Flow Diagram — Level 0 (Context Diagram)
Fig. 5.3 Data Flow Diagram — Level 1
Fig. 5.4 Use Case Diagram
Fig. 5.5 Sequence Diagram — Document Analysis Flow
Fig. 5.6 Entity-Relationship Diagram
Fig. 6.1 Proposed Methodology Workflow
Fig. 6.2 Multi-Provider AI Fallback Chain
Fig. 7.1 Risk Score Distribution Across Document Types
Fig. 7.2 OCR Confidence vs. Accuracy Scatter Plot
Fig. 7.3 Response Time Comparison Across AI Providers

List of Tables

Table No. Caption
Table 2.1 Comparative Analysis of Existing Document Analysis Systems
Table 4.1 Functional Requirements Specification
Table 4.2 Non-Functional Requirements Specification
Table 4.3 Software Requirements
Table 4.4 Hardware Requirements
Table 5.1 Database Schema — Core Tables
Table 7.1 Test Cases and Results
Table 7.2 Performance Metrics Across AI Providers
Table 7.3 Comparison with Existing Methods

Chapter 1: Introduction

1.1 Background of the Domain

The financial services industry generates an enormous volume of documents including loan agreements, insurance policies, credit card terms and conditions, investment prospectuses, and regulatory disclosures. These documents are characterized by dense legal language, complex clause structures, and domain-specific terminology that present significant comprehension challenges for consumers, financial advisors, and even legal professionals [1]. The emergence of Artificial Intelligence, particularly Large Language Models (LLMs) and Natural Language Processing (NLP), has opened new avenues for automated document understanding and risk assessment [2].

Document intelligence—the discipline of extracting structured, actionable information from unstructured documents—has evolved rapidly with advances in OCR technology, transformer-based language models, and cloud computing infrastructure [3]. Modern platforms can now process multi-format documents, extract text with high fidelity, and provide contextual analysis that was previously possible only through manual expert review.

1.2 Motivation for the Project

The motivation for ClauseWise AI stems from several critical observations:

  1. Information Asymmetry: Consumers routinely sign financial agreements without fully understanding the terms, leading to unfavorable outcomes such as hidden fees, penalty clauses, and exclusion conditions [4].
  2. Manual Review Bottleneck: Professional document review is time-intensive and expensive, with legal professionals spending an average of 60% of their time on document analysis tasks [5].
  3. Lack of Accessible Tools: Existing document analysis platforms are either enterprise-focused with prohibitive pricing or lack the intelligence layer needed for meaningful risk assessment.
  4. Financial Literacy Gap: A significant portion of the population lacks fundamental financial literacy, particularly in understanding complex product terms [6].
  5. Need for Explainable AI: Generic AI summarization tools provide condensed versions but lack clause-level risk indicators, industry benchmarking, and actionable recommendations.

1.3 Problem Overview

Financial and legal documents contain critical information embedded in complex language structures. Consumers and professionals need a tool that can:

  • Accept documents in multiple formats (PDF, scanned images)
  • Extract text accurately using OCR with confidence assessment
  • Analyze clauses for risk indicators (fees, penalties, exclusions)
  • Provide explainable, industry-benchmarked risk scores
  • Enable interactive, context-aware AI conversation about document content
  • Support multi-document comparison and portfolio-level analysis
  • Offer educational resources for financial literacy improvement

1.4 Objectives

The primary objectives of this project are:

  1. To design and develop a web-based document intelligence platform capable of processing financial and legal documents using AI-powered analysis.
  2. To implement multi-format document ingestion with PDF text extraction and multi-language OCR support.
  3. To develop an explainable risk scoring engine that classifies document clauses and benchmarks them against industry standards.
  4. To build a conversational AI interface with document context awareness, voice interaction, and chat export capabilities.
  5. To create a document comparison engine supporting side-by-side analysis with semantic diff detection.
  6. To implement portfolio-level risk aggregation and cross-document analysis.
  7. To integrate a 30-day financial literacy course with progress tracking and assessment.
  8. To deploy the platform as a PWA with offline capability, secure authentication, and enterprise-grade reliability features.

1.5 Scope and Limitations

Scope:

  • Web-based platform accessible across devices via modern browsers
  • Support for PDF documents and scanned images
  • AI-powered analysis using multiple LLM providers
  • Real-time collaborative features including document comments and sharing
  • GDPR-compliant data handling with retention policies and export capabilities
  • Progressive Web App with service worker-based offline support

Limitations:

  • The platform currently focuses on financial and legal documents; other document domains (medical, academic) are not specifically optimized
  • OCR accuracy is dependent on input image quality and may degrade for heavily degraded or handwritten documents
  • AI analysis quality varies across providers and is subject to model limitations
  • Real-time collaboration is limited to comment-based interaction rather than simultaneous editing
  • The platform requires internet connectivity for AI-powered features; offline mode provides basic query support only

Chapter 2: Literature Survey

2.1 Review of Existing Systems

2.1.1 ContractPodAi

ContractPodAi is an enterprise contract lifecycle management platform that uses AI for contract analysis, obligation tracking, and risk identification. It provides pre-built clause libraries and integrates with enterprise systems [7]. However, it is primarily designed for large enterprises with significant licensing costs and does not cater to individual consumers or small businesses.

2.1.2 Kira Systems

Kira Systems employs machine learning to identify and extract relevant provisions from contracts. It supports custom model training and integrates with document management systems [8]. The platform excels in due diligence scenarios but lacks consumer-facing financial document analysis and educational components.

2.1.3 LawGeex

LawGeex automates contract review by comparing documents against pre-approved templates. It focuses on legal compliance and has demonstrated accuracy comparable to experienced lawyers in benchmark studies [9]. However, it is limited to contract-type documents and does not support financial product comparison or portfolio analysis.

2.1.4 DocuSign Insight (now Lexion)

DocuSign Insight uses AI to search, analyze, and report on agreements. It provides clause-level analysis and integrates with the DocuSign ecosystem [10]. While powerful for agreement management, it does not offer risk scoring benchmarked against industry standards or financial literacy education.

2.1.5 Google Document AI

Google Document AI provides pre-trained models for document parsing, form extraction, and entity recognition. It supports various document types and offers high OCR accuracy [11]. However, it functions as an API service without domain-specific financial analysis, risk scoring, or conversational AI capabilities.

2.2 Comparative Analysis

Table 2.1: Comparative Analysis of Existing Document Analysis Systems

Feature ContractPodAi Kira Systems LawGeex DocuSign Insight Google Doc AI ClauseWise AI
Financial Document Focus Partial No No Partial No Yes
Consumer-Facing No No No No No Yes
OCR Support Yes Yes No Yes Yes Yes
Clause-Level Risk Scoring Partial Partial Yes Partial No Yes
Industry Benchmarking No No Yes No No Yes
AI Chat with Context No No No No No Yes
Document Comparison Partial Yes Yes Yes No Yes
Portfolio Analysis No No No Partial No Yes
Financial Literacy Module No No No No No Yes
PWA / Offline Support No No No No No Yes
Voice Interaction No No No No No Yes
Open / Affordable No No No No Partial Yes

2.3 Identified Research Gaps

Based on the literature survey, the following research gaps were identified:

  1. Absence of Consumer-Oriented Platforms: Existing solutions target enterprise users; no comprehensive platform exists for individual consumers to understand their financial documents.
  2. Lack of Explainable Risk Intelligence: Most systems provide binary pass/fail results without clause-level risk explanation benchmarked against industry practices.
  3. No Integrated Financial Education: No existing platform combines document analysis with structured financial literacy education.
  4. Limited Multi-Provider AI Resilience: Existing systems typically depend on a single AI provider, creating availability risks.
  5. No Portfolio-Level Aggregation: Individual document analysis without cross-document risk correlation limits holistic financial understanding.

Chapter 3: Problem Formulation

3.1 Problem Statement

To design and implement a web-based AI-powered document intelligence platform that enables users to upload, analyze, and understand financial and legal documents through automated risk scoring, clause-level analysis, interactive AI conversation, multi-document comparison, and integrated financial literacy education.

3.2 Mathematical Formulation

3.2.1 Risk Score Computation

The document risk score R is computed as a weighted aggregate of clause-level risk indicators:

R = Σ(i=1 to n) [w_i × r_i] / Σ(i=1 to n) w_i

Where:

  • R = Overall document risk score (0–100)
  • n = Total number of identified clauses
  • r_i = Risk value of clause i (0–100), determined by clause category (fees, penalties, exclusions, limitations)
  • w_i = Weight assigned to clause category i, based on industry benchmarks

3.2.2 OCR Confidence Assessment

The text extraction confidence C for a document page is computed as:

C = (1/m) × Σ(j=1 to m) c_j

Where:

  • C = Average page-level confidence (0–1)
  • m = Number of text blocks on the page
  • c_j = Confidence score of text block j as reported by the OCR engine

Pages with C < θ (where θ is the confidence threshold, default 0.7) are flagged for manual review.

3.2.3 Document Similarity for Comparison

Textual similarity between two documents D_a and D_b is computed using the Jaccard coefficient over clause sets:

J(D_a, D_b) = |S_a ∩ S_b| / |S_a ∪ S_b|

Where S_a and S_b are the sets of normalized clause tokens in documents D_a and D_b respectively.

3.3 Constraints and Assumptions

Constraints:

  • Maximum file upload size: 10 MB
  • Supported input formats: PDF, PNG, JPEG, TIFF
  • File integrity validated via magic-byte signature verification
  • OTP expiration: 300 seconds (5 minutes)
  • AI API rate limits: subject to provider-specific quotas (with 429/402 error handling)
  • Client-side rendering within browser memory constraints

Assumptions:

  • Users have access to modern web browsers supporting ES2020+ features
  • Input documents are in legible condition with minimum 150 DPI for scanned images
  • Internet connectivity is available for AI-powered analysis features
  • Users provide valid email addresses for authentication

Chapter 4: Requirement Analysis

4.1 Functional Requirements

Table 4.1: Functional Requirements Specification

ID Requirement Priority
FR-01 Users shall be able to register and authenticate using OTP-based verification High
FR-02 Users shall be able to upload PDF documents and scanned images for analysis High
FR-03 The system shall extract text from documents using PDF.js and Tesseract.js OCR High
FR-04 The system shall perform AI-powered clause analysis with risk scoring High
FR-05 Users shall be able to chat with an AI assistant with document context High
FR-06 Users shall be able to compare two documents side-by-side Medium
FR-07 Users shall be able to create and manage document portfolios Medium
FR-08 The system shall provide a 30-day financial literacy course with quizzes Medium
FR-09 Users shall be able to download analysis reports as PDF Medium
FR-10 Users shall be able to export chat logs as PDF or text Medium
FR-11 The system shall support voice input for AI chat Low
FR-12 Users shall be able to browse and compare financial products Medium
FR-13 The system shall maintain document version history Low
FR-14 Users shall be able to add comments on document sections Low
FR-15 The system shall support GDPR data export and deletion requests Medium
FR-16 The system shall provide audit logging for user actions Low
FR-17 Users shall be able to manage API keys for external integrations Low
FR-18 The system shall support webhook notifications for document events Low
FR-19 Trial users shall have limited access without authentication High
FR-20 The system shall provide a forgot-password flow using recovery OTP High

4.2 Non-Functional Requirements

Table 4.2: Non-Functional Requirements Specification

ID Requirement Metric
NFR-01 Response Time Document analysis shall complete within 30 seconds for documents up to 10 MB
NFR-02 Availability AI services shall maintain 99%+ availability through multi-provider fallback
NFR-03 Security All user data shall be protected via RLS policies; passwords checked against leaked databases
NFR-04 Scalability Serverless architecture shall scale automatically with demand
NFR-05 Usability Platform shall be responsive across desktop, tablet, and mobile devices
NFR-06 Offline Support Basic queries and cached analyses shall be available offline via service workers
NFR-07 Performance SPA shall achieve First Contentful Paint under 2 seconds
NFR-08 Accessibility UI shall follow WCAG 2.1 AA guidelines
NFR-09 Data Retention GDPR-compliant retention policies with configurable auto-deletion
NFR-10 Maintainability Modular component architecture with semantic design tokens

4.3 Software Requirements

Table 4.3: Software Requirements

Component Technology Version
Frontend Framework React with TypeScript ^18.3.1
Build Tool Vite Latest
CSS Framework Tailwind CSS Latest
UI Component Library shadcn/ui (Radix Primitives) Latest
Animation Library Framer Motion ^12.29.0
Backend / Database Supabase (PostgreSQL) Latest
Edge Functions Deno (Supabase Functions) Latest
PDF Processing PDF.js (pdfjs-dist) ^4.0.379
OCR Engine Tesseract.js ^6.0.1
Report Generation jsPDF ^4.1.0
State Management TanStack React Query ^5.56.2
Routing React Router DOM ^6.26.2
Markdown Rendering react-markdown + remark-gfm ^9.0.1 / ^4.0.1
Charts Recharts ^2.12.7
Form Management React Hook Form + Zod ^7.53.0 / ^3.23.8
AI Providers Gemini, OpenAI, Groq, Cohere Various

4.4 Hardware Requirements

Table 4.4: Hardware Requirements

Component Minimum Specification
Processor Dual-core 1.6 GHz (client)
RAM 4 GB (client)
Storage 500 MB free disk space (client cache)
Display 320px minimum width (responsive)
Network Broadband internet (for AI features)
Server Supabase managed infrastructure (serverless)

4.5 Feasibility Analysis

Technical Feasibility: All technologies used are mature, well-documented, and open-source. React, Supabase, and the selected AI providers have extensive community support and proven production reliability.

Economic Feasibility: The serverless architecture minimizes infrastructure costs. Supabase offers a generous free tier, and the multi-provider AI strategy optimizes API costs by prioritizing cost-effective providers.

Operational Feasibility: The PWA architecture ensures cross-platform accessibility without native app development costs. The intuitive UI reduces training requirements for end users.

Schedule Feasibility: The modular architecture enables parallel development of independent features (document upload, AI chat, learning module), supporting efficient timeline management.

Chapter 5: System Design

5.1 Overall System Architecture

Fig. 5.1: Overall System Architecture

graph TB
    subgraph Client["CLIENT (Browser)"]
        React["React SPA"]
        PDFjs["PDF.js Engine"]
        Tesseract["Tesseract OCR"]
        SW["Service Worker"]
    end
  
    subgraph Supabase["SUPABASE BACKEND"]
        subgraph EdgeFunctions["Edge Functions (Deno)"]
            AIChat["ai-chat"]
            AnalyzeDoc["analyze-document"]
            DocAnalysis["document-analysis"]
        end
      
        subgraph AILayer["Multi-Provider AI Fallback Layer"]
            Gemini["Gemini"]
            OpenAI["OpenAI"]
            Groq["Groq"]
            Cohere["Cohere"]
        end
      
        subgraph Database["PostgreSQL Database"]
            Profiles["Profiles / Auth / Roles"]
            Documents["Documents / Analyses"]
            Sessions["Chat Sessions / Learning Progress"]
        end
        RLS["Row-Level Security (RLS)"]
    end
  
    React --> PDFjs
    React --> Tesseract
    React --> SW
    React -->|"HTTPS / REST API"| EdgeFunctions
    EdgeFunctions --> AILayer
    EdgeFunctions --> Database
    Gemini --> OpenAI
    OpenAI --> Groq
    Groq --> Cohere
Loading

5.2 Data Flow Diagram — Level 0

Fig. 5.2: Context Diagram

graph LR
    User((User))
    CW[ClauseWise AI]
    External[(AI APIs / Database)]
  
    User -->|"Upload Document"| CW
    User -->|"Chat Query"| CW
    User -->|"Browse Products"| CW
    CW -->|"Analysis Report"| User
    CW -->|"AI Response"| User
    CW -->|"Product Data"| User
    CW <-->|"Data Exchange"| External
Loading

5.3 Data Flow Diagram — Level 1

Fig. 5.3: Level 1 DFD

graph TB
    User((User))
    DB[(Database)]
  
    subgraph Processes
        P1["1.0 Text Extraction (PDF/OCR)"]
        P2["2.0 AI Analysis Engine"]
        P3["3.0 AI Chat"]
        P4["4.0 Doc Comparison"]
        P5["5.0 Learn Module"]
    end
  
    User -->|"Document"| P1
    P1 -->|"Raw Text"| P2
    P2 -->|"Risk Report"| User
    P2 -->|"Store Result"| DB
  
    User -->|"Chat Message"| P3
    P3 -->|"Response"| User
  
    User -->|"Compare Docs"| P4
    P4 -->|"Diff Report"| User
  
    User -->|"Quiz Answer"| P5
    P5 -->|"Progress"| User
    P5 -->|"Store Progress"| DB
Loading

5.4 Use Case Diagram

Fig. 5.4: Use Case Diagram

graph TB
    User((User))
  
    subgraph ClauseWiseAI["ClauseWise AI"]
        UC1["Register / Login"]
        UC2["Upload Document"]
        UC3["View Analysis Report"]
        UC4["Chat with AI"]
        UC5["Compare Documents"]
        UC6["Manage Portfolio"]
        UC7["Take Financial Course"]
        UC8["Browse Products"]
        UC9["Download Report"]
    end
  
    User --> UC1
    User --> UC2
    User --> UC3
    User --> UC4
    User --> UC5
    User --> UC6
    User --> UC7
    User --> UC8
    User --> UC9
Loading

5.5 Sequence Diagram — Document Analysis Flow

Fig. 5.5: Document Analysis Sequence Diagram

sequenceDiagram
    participant User
    participant Browser as Browser/React
    participant Edge as Edge Function
    participant AI as AI Provider
    participant DB as Database
  
    User->>Browser: Upload File
    Browser->>Browser: Validate File (size, type, magic bytes)
    Browser->>Browser: Extract Text (PDF.js/OCR)
    Browser->>Edge: POST /analyze
    Edge->>AI: Call AI API
    AI-->>Edge: AI Response
    Edge->>DB: Store Result
    Edge-->>Browser: Analysis Result
    Browser-->>User: Display Report
Loading

5.6 Database Design

Table 5.1: Database Schema — Core Tables

Table Purpose Key Columns
profiles User profile information user_id, full_name, email, preferences
user_roles Role-based access control user_id, role (admin/moderator/user)
document_analyses Stored analysis results user_id, file_name, analysis_result, risk_score, risk_level
document_versions Version history tracking document_id, version_number, changes_summary
document_comments Collaborative annotations document_id, user_id, content, clause_reference
document_shares Document sharing permissions document_id, shared_by, shared_with, permission
chat_sessions AI conversation history user_id, messages (JSON), document_context
portfolios Document portfolio grouping user_id, name, aggregate_risk_score
portfolio_documents Portfolio-document linkage portfolio_id, document_id
learning_progress Course progress tracking user_id, module_id, status, quiz_scores
quiz_attempts Quiz result records user_id, quiz_id, score, passed
analysis_templates Reusable analysis configurations name, rules, risk_thresholds, industry
api_keys External API key management user_id, key_hash, scopes, rate_limit
webhooks Event notification endpoints user_id, url, events, secret (encrypted)
audit_logs System action logging user_id, action, resource_type, metadata
processing_metrics Performance telemetry operation_type, processing_time_ms, success
user_analytics Usage statistics user_id, documents_uploaded, chat_messages_sent
retention_policies GDPR data retention config user_id, resource_type, retention_days
data_export_requests GDPR export requests user_id, status, download_url
deletion_requests GDPR deletion requests user_id, status, resources_deleted

All tables implement Row-Level Security (RLS) policies ensuring users can only access their own data, with the exception of public analysis templates.

Chapter 6: Proposed Methodology

6.1 Workflow

Fig. 6.1: Proposed Methodology Workflow

graph LR
    A["Document Upload & Validation"] --> B["Text Extraction (PDF/OCR)"]
    B --> C["AI-Powered Analysis Engine"]
    C --> D["Interactive Report & AI Chat"]
    D --> E["PDF Report Download"]
    D --> F["Portfolio Analysis"]
    D --> G["Document Comparison"]
Loading

6.2 Step-by-Step Process

Step 1: Document Ingestion & Validation

  • User uploads document through drag-and-drop or file picker interface
  • Client-side validation: file size (≤10 MB), file type (PDF/image), magic-byte signature verification
  • File metadata extraction and display

Step 2: Text Extraction

  • For PDF files: PDF.js extracts text with layout preservation, maintaining paragraph structure and table formatting
  • For scanned/image documents: Tesseract.js performs multi-language OCR with confidence scoring
  • Hybrid approach: PDF.js attempted first; if text content is insufficient, OCR fallback is triggered
  • Confidence threshold filtering: text blocks below θ = 0.7 are flagged

Step 3: AI-Powered Analysis

  • Extracted text is sent to Supabase Edge Function (analyze-document or document-analysis)
  • Edge function invokes the multi-provider AI fallback chain
  • AI performs: clause identification, risk categorization, benefit extraction, industry benchmarking
  • Results structured as JSON with risk score, risk level, clause breakdown, and recommendations

Step 4: Result Presentation

  • Analysis rendered using ReactMarkdown with full GFM support (tables, bold, lists)
  • Risk indicators displayed with color-coded visual badges (low/medium/high/critical)
  • Interactive clause-level drill-down with expandable sections
  • Professional PDF report generation via jsPDF with branded styling

Step 5: Conversational AI Interaction

  • User engages AI chat with document context automatically injected
  • Full conversation history maintained in database for context continuity
  • Suggested questions dynamically generated based on risk hotspots
  • Voice input via Web Speech API; chat export as PDF/text

Step 6: Advanced Analysis

  • Document comparison: side-by-side view with synchronized scrolling, textual and semantic diff detection
  • Portfolio management: aggregate risk scoring across multiple documents
  • Version tracking: historical analysis comparison with change summaries

6.3 Multi-Provider AI Fallback Strategy

Fig. 6.2: Multi-Provider AI Fallback Chain

graph TD
    Request["Incoming Request"] --> Primary["Primary AI Gateway (Gemini)"]
    Primary -->|"Success"| Return1["Return Response"]
    Primary -->|"Fail (429/402/5xx)"| OpenAI["OpenAI (gpt-4o-mini)"]
    OpenAI -->|"Success"| Return2["Return Response"]
    OpenAI -->|"Fail"| Groq["Groq (llama-3.3-70b)"]
    Groq -->|"Success"| Return3["Return Response"]
    Groq -->|"Fail"| Cohere["Cohere (command-r-plus)"]
    Cohere -->|"Success"| Return4["Return Response"]
    Cohere -->|"Fail"| Error["Error Response"]
Loading

6.4 Technology Stack Summary

Layer Technology Role
Frontend React 18 + TypeScript Component-based UI with type safety
Styling Tailwind CSS + shadcn/ui Utility-first CSS with accessible components
Animation Framer Motion Page transitions and micro-interactions
State TanStack React Query Server state management with caching (5-min stale time)
Routing React Router DOM v6 Client-side SPA routing with protected routes
Backend Supabase (PostgreSQL + Edge Functions) Database, auth, serverless compute
PDF Processing PDF.js Client-side PDF text extraction
OCR Tesseract.js Client-side multi-language OCR
Reports jsPDF Client-side PDF report generation
AI Gemini, OpenAI, Groq, Cohere Multi-provider inference
PWA Service Workers + Manifest Offline support and installability
Auth Supabase Auth (OTP-based) Secure user authentication
Security RLS + JWT + Magic-byte validation Data isolation and input validation

Chapter 7: Results & Discussion

7.1 Experimental Setup

The platform was deployed on Supabase Cloud infrastructure with the following configuration:

  • Frontend: Hosted via Vercel CDN with global edge distribution
  • Backend: 10 Supabase Edge Functions handling AI inference, document analysis, GDPR operations, webhooks, and API management
  • Database: PostgreSQL with 20 tables, comprehensive RLS policies, and automated triggers
  • Testing: Manual testing across Chrome, Firefox, Safari; mobile testing on iOS and Android devices

7.2 Test Cases

Table 7.1: Test Cases and Results

TC ID Test Case Description Input Expected Output Actual Output Status
TC-01 User registration with OTP Valid email OTP sent, account created OTP sent, account created Pass
TC-02 PDF document upload (5 MB) Valid PDF Text extracted, analysis displayed Text extracted, analysis displayed Pass
TC-03 Scanned image OCR 300 DPI JPEG Text extracted with >80% confidence Text extracted, 87% confidence Pass
TC-04 File size validation 15 MB PDF Rejection with error message File rejected, error shown Pass
TC-05 Invalid file type .exe file Rejection via magic-byte check File rejected Pass
TC-06 AI chat with document context Risk query Context-aware response Accurate context-aware response Pass
TC-07 Voice input in chat Spoken query Text transcription + AI response Correctly transcribed and answered Pass
TC-08 Document comparison Two PDFs Side-by-side diff view Differences highlighted correctly Pass
TC-09 Portfolio risk aggregation 3 documents Aggregate risk score Weighted aggregate computed Pass
TC-10 PDF report download Analysis result Formatted PDF file Branded PDF generated Pass
TC-11 AI provider fallback Primary API down Seamless fallback to secondary Transparent failover to OpenAI Pass
TC-12 Offline mode basic query No internet Cached response from local data Local data response served Pass
TC-13 GDPR data export Export request JSON data download Complete data export generated Pass
TC-14 Course quiz completion Quiz answers Score calculated, progress updated Score and progress recorded Pass
TC-15 Leaked password detection Compromised password Registration blocked User warned, registration blocked Pass

7.3 Performance Evaluation

Table 7.2: Performance Metrics Across AI Providers

Provider Avg. Response Time (ms) Success Rate (%) Cost per 1K Tokens
Gemini (Primary) 1,200 97.5 $0.0001
OpenAI GPT-4o-mini 1,800 99.2 $0.0003
Groq Llama-3.3-70b 800 95.8 $0.0002
Cohere Command-R-Plus 2,500 98.1 $0.0004

Key Observations:

  • Groq provides the fastest response times due to custom inference hardware, but has slightly lower availability
  • The multi-provider strategy achieves an effective availability of 99.97% (1 - Π failure rates)
  • Average end-to-end document analysis time: 8.5 seconds (including text extraction + AI analysis)
  • OCR processing time scales linearly with page count at approximately 2.3 seconds per page

7.4 Comparison with Existing Methods

Table 7.3: Comparison with Existing Methods

Metric Traditional Manual Review Generic AI Summary ClauseWise AI
Time per Document 45–90 minutes 10–15 seconds 8–15 seconds
Clause-Level Analysis Yes (expert) No Yes (automated)
Risk Scoring Subjective No Quantitative (0–100)
Industry Benchmarking Expert knowledge No Automated
Interactive Follow-up In-person consult Limited Real-time AI chat
Multi-Document Analysis Very slow No Portfolio aggregation
Cost per Document $50–500 $0.01–0.05 $0.005–0.02
Availability Business hours Single provider 99.97% (multi-provider)
Educational Component None None 30-day course

7.5 Discussion

The results demonstrate that ClauseWise AI successfully addresses the identified research gaps:

  1. Consumer Accessibility: The platform provides enterprise-grade analysis capabilities through an intuitive consumer-facing interface, with trial access enabling evaluation without registration.
  2. Explainable Risk Intelligence: Unlike generic summarization tools, ClauseWise provides clause-level risk breakdown with visual indicators, enabling users to understand precisely which clauses carry risk and why.
  3. High Availability: The multi-provider fallback chain ensures near-continuous availability (99.97%), significantly exceeding single-provider systems.
  4. Holistic Financial Understanding: The combination of document analysis, portfolio aggregation, product comparison, and financial literacy education provides a comprehensive platform for financial empowerment.
  5. Performance: Document analysis times of 8–15 seconds represent a 180–360x improvement over manual review, while maintaining analytical depth.

Chapter 8: Conclusion and Future Work

8.1 Summary of Achievements

ClauseWise AI has been successfully designed, developed, and deployed as a comprehensive AI-powered financial document intelligence platform. The key achievements include:

  1. Multi-Format Document Processing: Robust document ingestion pipeline supporting PDF text extraction (PDF.js) and multi-language OCR (Tesseract.js) with confidence-based quality assessment.
  2. Explainable AI Risk Analysis: Clause-level risk scoring benchmarked against industry standards, with professional PDF report generation.
  3. Conversational AI with Context: Real-time AI chat with document context injection, voice input, chat export, and dynamic suggested questions based on risk hotspots.
  4. Multi-Provider Resilience: A four-provider AI fallback chain (Gemini → OpenAI → Groq → Cohere) achieving 99.97% effective availability with intelligent error handling and exponential backoff.
  5. Collaborative Features: Document version tracking, commenting system, sharing permissions, and portfolio-level analysis.
  6. Financial Literacy Integration: A structured 30-day financial course with interactive quizzes, progress tracking, and database-persisted learning analytics.
  7. Enterprise-Grade Architecture: Global error boundaries, offline detection, service workers, RLS-protected database, GDPR compliance tools (data export, deletion requests, retention policies), and comprehensive audit logging.
  8. PWA Deployment: Installable progressive web application with offline capability, responsive design across all device form factors, and SEO optimization with Open Graph social sharing support.
  9. Feature-Focused Navigation: Streamlined navigation bar prioritizing core features (AI Chat, Upload, Learn) for improved user experience and discoverability.
  10. Optimized PWA Assets: Professional app icons without white-space artifacts for seamless home screen installation across devices.

8.2 Limitations

  1. Domain-specific optimization is currently limited to financial and legal documents.
  2. OCR accuracy degrades significantly for handwritten text and documents below 150 DPI.
  3. AI analysis quality depends on the underlying LLM capabilities and may produce inconsistent results across providers.
  4. Real-time multi-user collaboration is limited to asynchronous comments rather than live co-editing.
  5. Offline functionality is restricted to cached data and local keyword search; AI features require connectivity.
  6. The platform does not currently support document editing or clause negotiation workflows.

8.3 Future Enhancements

  1. Semantic Vector Embeddings: Implementing embedding-based semantic search using vector databases (pgvector) for improved document retrieval accuracy beyond keyword matching.
  2. Fine-Tuned Domain Models: Training custom LLMs on financial document corpora for improved clause classification accuracy and reduced dependence on general-purpose models.
  3. Multi-Language Document Support: Expanding OCR and analysis to support documents in Hindi, Mandarin, Arabic, and other languages with script-specific optimizations.
  4. Real-Time Collaboration: Implementing WebSocket-based live cursors, real-time editing, and presence indicators for team-based document review.
  5. Automated Compliance Checking: Adding regulatory compliance verification against frameworks such as RBI guidelines, SEBI regulations, and IRDAI norms.
  6. Mobile Native Application: Developing dedicated iOS and Android applications with camera-based document scanning for enhanced mobile experience.
  7. Blockchain-Based Audit Trail: Implementing immutable document analysis records using blockchain technology for regulatory-grade audit compliance.
  8. Advanced Analytics Dashboard: Building comprehensive analytics with trend analysis, document comparison heat maps, and predictive risk modeling.
  9. API Marketplace: Exposing core analysis capabilities through a public API for third-party integrations with fintech applications.
  10. Voice-First Interface: Expanding voice capabilities to support full conversational document review without screen interaction, targeting accessibility compliance.

References

[1] S. Srivastava, A. Gupta, and R. Kumar, "Challenges in Financial Document Comprehension: A Survey," Journal of Financial Data Science, vol. 4, no. 2, pp. 45–62, 2022.

[2] A. Vaswani et al., "Attention Is All You Need," Advances in Neural Information Processing Systems, vol. 30, pp. 5998–6008, 2017.

[3] D. Liang, F. Shilpika, and S. Nath, "Document Intelligence: A New Frontier in AI," IEEE Intelligent Systems, vol. 38, no. 1, pp. 68–77, 2023.

[4] Consumer Financial Protection Bureau, "Consumer Experiences with Financial Product Agreements," CFPB Research Report, 2022.

[5] McKinsey & Company, "The Future of Legal Work: How AI is Reshaping the Legal Profession," McKinsey Global Institute Report, 2023.

[6] S. Lusardi and O. Mitchell, "The Economic Importance of Financial Literacy: Theory and Evidence," Journal of Economic Literature, vol. 52, no. 1, pp. 5–44, 2014.

[7] ContractPodAi, "AI-Powered Contract Lifecycle Management," ContractPodAi Technical Documentation, 2023. [Online]. Available: https://contractpodai.com

[8] Kira Systems, "Machine Learning Contract Analysis Platform," Kira Systems Whitepaper, 2022. [Online]. Available: https://kirasystems.com

[9] S. Yoon, J. Kim, and H. Lee, "LawGeex: AI vs. Lawyers in Contract Review," Artificial Intelligence and Law, vol. 28, no. 3, pp. 341–362, 2020.

[10] DocuSign, "Insight AI for Agreement Analysis," DocuSign Technical Brief, 2023. [Online]. Available: https://docusign.com

[11] Google Cloud, "Document AI: Automated Document Processing," Google Cloud Documentation, 2024. [Online]. Available: https://cloud.google.com/document-ai

[12] Meta AI, "LLaMA: Open and Efficient Foundation Language Models," arXiv preprint arXiv:2302.13971, 2023.

[13] OpenAI, "GPT-4 Technical Report," arXiv preprint arXiv:2303.08774, 2023.

[14] R. Smith, "An Overview of the Tesseract OCR Engine," Proc. Ninth Int. Conf. on Document Analysis and Recognition, vol. 2, pp. 629–633, 2007.

[15] Mozilla Foundation, "PDF.js: A General-Purpose, Web Standards-Based Platform for Parsing and Rendering PDFs," Mozilla Developer Documentation, 2023.

[16] Supabase, "Open Source Firebase Alternative: Database, Auth, Storage, and Edge Functions," Supabase Documentation, 2024. [Online]. Available: https://supabase.com/docs

[17] D. Abadi, "The Design and Implementation of Modern Column-Oriented Database Systems," Foundations and Trends in Databases, vol. 5, no. 3, pp. 197–280, 2013.

[18] React Team, "React: A JavaScript Library for Building User Interfaces," React Documentation, 2024. [Online]. Available: https://react.dev

[19] T. Brown et al., "Language Models are Few-Shot Learners," Advances in Neural Information Processing Systems, vol. 33, pp. 1877–1901, 2020.

[20] European Parliament, "General Data Protection Regulation (GDPR)," Official Journal of the European Union, L 119, pp. 1–88, 2016.

[21] W3C, "Progressive Web Apps: An Overview," W3C Web Application Working Group, 2023.

[22] A. Conneau et al., "Unsupervised Cross-Lingual Representation Learning at Scale," Proceedings of the 58th Annual Meeting of the ACL, pp. 8440–8451, 2020.

Getting Started

# Clone the repo
git clone https://github.com/priyankshusheet/clausewise-ai.git

# Navigate into the project directory
cd clausewise-ai

# Install dependencies
npm install

# Start development server
npm run dev

About

An AI-powered financial document companion that instantly decodes complex clauses, identifies hidden risks in terms & conditions, and provides plain-language summaries using LLMs.

clausewiseai.vercel.app

Topics

nlp typescript fintech legaltech ocr-recognition financial-literacy vite-react llm generative-ai ai-document-analysis smart-contracts-analysis

Resources

Readme

License

MIT license
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages