🏛️ CCMS AI — Full Stack Hackathon Build Prompt SYSTEM OVERVIEW Build a production-ready full-stack web application called "JudgeAI — Court Judgment Intelligence System" for the Centre for e-Governance. The system ingests court judgment PDFs, uses AI to extract structured legal data, generates actionable compliance plans, routes them through a human verification workflow, and displays only verified data on a government decision-maker dashboard. TECH STACK Frontend: React (Vite) + Tailwind CSS + shadcn/ui Backend: Node.js + Express.js (REST API) Database: PostgreSQL (via Prisma ORM) AI Layer: Anthropic Claude API (claude-sonnet-4-20250514) with PDF document support PDF Handling: pdf-parse (text PDFs) + Tesseract.js (scanned/OCR PDFs) Auth: JWT-based role authentication (Uploader / Reviewer / Dashboard Viewer) File Storage: Local multer storage (or S3-compatible) Charts: Recharts Deployment: Single monorepo with /client and /server folders DATABASE SCHEMA (Prisma) Design the following models: User { id, name, email, passwordHash, role: UPLOADER|REVIEWER|VIEWER, department, createdAt } Judgment { id, caseNumber, courtName, pdfPath, uploadedBy, uploadedAt, status: PENDING|PROCESSING|EXTRACTED|VERIFIED|REJECTED } Extraction { id, judgmentId, caseTitle, caseNumber, courtName, dateOfOrder, bench, petitioner, respondent, petitionerAdvocate, respondentAdvocate, keyDirections: JSON[], // array of { directive, page, confidenceScore } timelines: JSON[], // array of { event, date, isInferred } relevantActs: string[], summaryText, rawExtractedText, aiConfidenceScore, extractionModel, extractedAt } ActionPlan { id, judgmentId, extractionId, complianceRequired: bool, appealConsideration: bool, appealLimitationDays: int, appealDeadline: date, responsibleDepartments: string[], priorityLevel: LOW|MEDIUM|HIGH|CRITICAL, actionItems: JSON[], // array of { action, owner, dueDate, isInferred } natureOfAction, aiRationale, generatedAt } VerificationRecord { id, judgmentId, extractionId, actionPlanId, reviewedBy, reviewedAt, status: APPROVED|EDITED|REJECTED, reviewerNotes, editedExtraction: JSON, // stores reviewer overrides editedActionPlan: JSON } DashboardEntry { id, verificationId, judgmentId, department, caseNumber, caseTitle, priorityLevel, complianceRequired, appealConsidering, keyActions: JSON[], importantDates: JSON[], status: ACTIVE|COMPLIED|APPEALED|CLOSED, createdAt, updatedAt } BACKEND API ROUTES Auth POST /api/auth/register — register with role POST /api/auth/login — returns JWT Judgment Upload & Processing POST /api/judgments/upload — multipart PDF upload, saves file, creates Judgment record, triggers async processing pipeline GET /api/judgments — list all judgments with status GET /api/judgments/:id — single judgment detail GET /api/judgments/:id/pdf — serve raw PDF for inline viewer AI Processing Pipeline (internal, triggered on upload) Build a pipeline service processingPipeline.js: Step 1 — PDF Text Extraction Use pdf-parse for digital PDFs Fall back to Tesseract.js OCR for scanned PDFs Detect which mode was used, store in metadata Step 2 — AI Extraction via Claude API Send the full extracted text (chunked if >100k chars) to Claude with this system prompt: You are a legal document analysis AI for the Indian government's Court Case Monitoring System. Analyze the court judgment text provided and extract structured data. Return ONLY valid JSON with this exact schema: { "caseTitle": "", "caseNumber": "", "courtName": "", "dateOfOrder": "YYYY-MM-DD or null", "bench": [], "petitioner": { "name": "", "advocate": "" }, "respondent": { "name": "", "advocate": "" }, "keyDirections": [ { "directive": "", "pageHint": "", "confidenceScore": 0.0-1.0, "isExplicit": true/false } ], "timelines": [ { "event": "", "date": "YYYY-MM-DD or null", "isInferred": true/false, "inferenceReason": "" } ], "relevantActs": [], "summaryText": "", "overallConfidenceScore": 0.0-1.0, "confidenceNotes": "" } Step 3 — Action Plan Generation via second Claude call Using the extraction JSON, call Claude again with: Based on this court judgment extraction for an Indian government department, generate an actionable compliance plan. Return ONLY valid JSON: { "complianceRequired": true/false, "complianceRationale": "", "appealConsideration": true/false, "appealRationale": "", "appealLimitationDays": number or null, "appealDeadline": "YYYY-MM-DD or null", "priorityLevel": "LOW|MEDIUM|HIGH|CRITICAL", "priorityRationale": "", "responsibleDepartments": [], "natureOfAction": "", "actionItems": [ { "action": "", "owner": "", "dueDate": "YYYY-MM-DD or null", "isInferred": true/false, "urgency": "LOW|MEDIUM|HIGH" } ], "aiRationale": "" } Update Judgment status throughout: PROCESSING → EXTRACTED Verification GET /api/verify/queue — list all EXTRACTED judgments pending review GET /api/verify/:judgmentId — get extraction + action plan side-by-side with PDF URL POST /api/verify/:judgmentId/approve — approve as-is, create DashboardEntry, mark VERIFIED POST /api/verify/:judgmentId/edit — body contains { editedExtraction, editedActionPlan, reviewerNotes }, saves edits, creates DashboardEntry POST /api/verify/:judgmentId/reject — body contains { reviewerNotes }, marks REJECTED Dashboard GET /api/dashboard/entries — all verified dashboard entries, supports query params: ?department=&priority=&status=&dateFrom=&dateTo= GET /api/dashboard/stats — { total, byPriority, byDepartment, byStatus, upcomingDeadlines[] } GET /api/dashboard/entries/:id — single entry detail PATCH /api/dashboard/entries/:id/status — update compliance status FRONTEND PAGES & COMPONENTS 1. Login Page (/login) Clean government-style login. Role-based redirect after login. 2. Upload Page (/upload) — UPLOADER role Drag-and-drop PDF upload zone with file validation Shows upload progress bar After upload: shows real-time processing status ticker: Uploading → Extracting Text → Running AI Analysis → Generating Action Plan → Ready for Review Use polling (GET /api/judgments/:id) every 3 seconds to update status 3. Verification Queue (/verify) — REVIEWER role Table of all EXTRACTED judgments: Case No. | Court | Date | AI Confidence | Actions Color-coded confidence badges: Green (>0.85) / Yellow (0.6–0.85) / Red (<0.6) Click row → Verification Detail Page 4. Verification Detail Page (/verify/:id) — KEY PAGE Split-panel layout: Left Panel — PDF Viewer Render PDF inline using react-pdf (PDF.js) Page navigation controls Right Panel — AI Extraction Review Tabbed interface: Tab 1: Extracted Data Show each field with its value AND confidence score Low-confidence fields highlighted in yellow with warning icon Every field is editable inline (click to edit) Show AI Confidence: 87% badge at top Tab 2: Action Plan Display all action items in card format Each item shows: Action | Owner | Due Date | Urgency badge | Inferred? badge All fields editable Tab 3: Review Notes Textarea for reviewer notes Mandatory if rejecting Bottom Action Bar: ✅ Approve (green) | ✏️ Approve with Edits (blue) | ❌ Reject (red) Confirmation modal before any action 5. Dashboard (/dashboard) — VIEWER + all roles Header Stats Row: [ Total Cases ] [ Pending Action ] [ High Priority ] [ Upcoming Deadlines ] Filter Bar: Department | Priority | Status | Date Range | Search Main Content: Left: Data table of all verified entries with sortable columns Clicking a row opens a slide-over detail panel (not a new page) showing full action plan Charts Section (below table): Bar chart: Cases by Department Donut chart: Priority Distribution Timeline chart: Deadlines in next 30 days (use Recharts) Individual Case Card View (toggle): Each card shows: ┌─────────────────────────────────────────┐ │ [CRITICAL] WP/1234/2024 │ │ Petitioner vs State of Karnataka │ │ Karnataka High Court | 12 Mar 2025 │ ├─────────────────────────────────────────┤ │ 🏛 Dept: Revenue Department │ │ ⚡ Action: Comply with directive │ │ 📅 Deadline: 15 Jun 2025 (43 days) │ │ 🔔 Appeal: Under Consideration │ └─────────────────────────────────────────┘ 6. Shared Components <Navbar> with role-aware nav links + logout <ConfidenceBadge score={0.87} /> — colored pill <PriorityBadge level="HIGH" /> — colored with icon <StatusTimeline steps={[...]} current="EXTRACTED" /> — horizontal stepper <DeadlineCountdown date="2025-06-15" /> — shows days remaining, red if <7 days ROLE-BASED ACCESS CONTROL UPLOADER → can access: /upload, /dashboard (view only) REVIEWER → can access: /verify, /verify/:id, /dashboard (view only) VIEWER → can access: /dashboard only Protect routes both on frontend (redirect) and backend (JWT middleware + role check). AI PROMPT ENGINEERING DETAILS For the extraction call, prepend this context to improve Indian legal document handling: This is an Indian High Court judgment. Common patterns: - Case numbers: W.P., W.A., O.S., Crl., CMA formats - Dates in DD.MM.YYYY or DD/MM/YYYY format - Directives often preceded by "It is hereby directed", "The respondent shall", "Liberty is granted" - Limitation periods under Limitation Act 1963: typically 90 days for High Court appeals - Government respondents are often referred to as "State", "Union of India", department names Extract with high precision. If unsure, mark confidenceScore below 0.7 and explain in confidenceNotes. SAMPLE DATA & SEEDING Create a seed.js script that: Creates 3 demo users (one per role) Inserts 5 sample judgment records in various statuses Inserts corresponding mock extractions and action plans Creates 3 verified dashboard entries UI DESIGN SYSTEM Use a professional government-appropriate design: Primary color: #1a3c6e (deep navy blue) Accent: #f59e0b (amber — for warnings/deadlines) Success: #16a34a, Danger: #dc2626 Font: Inter (clean, readable) All tables must have zebra striping, hover states, and sticky headers Mobile responsive (government officials use tablets) ERROR HANDLING & EDGE CASES If Claude API fails: mark judgment as EXTRACTION_FAILED, show retry button If PDF is scanned and OCR confidence <50%: warn reviewer with banner "Low quality scan — verify carefully" If no date of order found: flag as dateOfOrder: null and highlight in red on verification screen API rate limiting: 429 handler with exponential backoff retry (max 3 attempts) File size limit: reject PDFs >50MB with clear error message DELIVERABLE STRUCTURE / ├── client/ # React + Vite frontend │ ├── src/ │ │ ├── pages/ │ │ │ ├── Login.jsx │ │ │ ├── Upload.jsx │ │ │ ├── VerifyQueue.jsx │ │ │ ├── VerifyDetail.jsx │ │ │ └── Dashboard.jsx │ │ ├── components/ │ │ └── api/ # axios service layer ├── server/ # Node.js + Express backend │ ├── routes/ │ ├── services/ │ │ ├── pdfExtractor.js │ │ ├── claudeExtractor.js │ │ ├── actionPlanGenerator.js │ │ └── processingPipeline.js │ ├── prisma/schema.prisma │ └── seed.js └── README.md JUDGING CRITERIA ALIGNMENT CriterionImplementationAccuracy of extractionConfidence scores, OCR fallback, Claude with legal contextAction plan qualityStructured JSON with rationale, inferred vs explicit flagsHuman verification UXSplit-panel PDF+AI view, inline editing, confidence highlightsDashboard clarityFilters, charts, deadline countdowns, department viewsExplainabilityEvery field shows AI rationale, confidence score, and source hint