Traverse

There are 8.3 billion people in the world today. They speak over 7,000 languages, live across 195 countries — and in 2025 alone, 1.5 billion of them crossed an international border.

But the world is getting harder to move through. Across the roughly 38,000 travel corridors worldwide, each has its own rules, its own required documents, and its own language. As many as one in four visa applications get refused — and the most common reasons are preventable. Incomplete forms. Missing documents. Contradictions the applicant never noticed.

Traverse was built to fix that.

Three AI agents review your visa application the way an experienced immigration team would. Traverse researches the travel requirements for any corridor, reads every document — in any of 40 languages — and tells you exactly what to fix before you submit. The Research, Document Intelligence, and Advisory agents coordinate to catch the errors that cause preventable rejections — across documents, across languages, for any corridor in the world.

Built for an Anthropic hackathon — showcasing extended thinking, web search, vision, and multi-agent coordination.

Quick Start

Get an Anthropic API key
Clone and install: npm install
Add API key to .env.local
Run: npm run dev
Open http://localhost:3000
Click a demo persona in the sidebar or enter your own travel details

How It Looks

As many as one in four visa applications get refused — and the most common reasons are preventable. Traverse uses agentic AI to provide a solution to a complex situation: there are over 195 countries in the world, resulting in 37,830 global travel corridors (where a travel corridor refers to 1-way direction from country A to country B, ex: travel from USA to France), and the thousands of languages adds complexity in understanding what travel documents are needed as global tensions mount.

For the demo, we use synthetic data across three fictional profiles, including Amara from Nigeria who needs a student visa for her master's program in London; and Carlos, a freelance photographer from Brazil who needs to check travel documents for his 3 week trip to Japan.

Priya is a software engineer in Bangalore applying for a business visa to Germany. She has nine documents, some in Hindi, some in English, and one contains a contradiction she doesn't know about.

Her documents include this bilingual employment certificate. The Hindi section says she's a permanent employee earning ₹85,000/month. The English cover letter will say something different — and she doesn't know.

She clicks Check Requirements. The analysis page greets her in English, Hindi, and German — matching her corridor — and the Research Agent starts searching live government sources.

Within seconds, the Research Agent returns visa type, fees, processing time, and 11 structured requirements, sourced from the German Embassy and VFS Global.

Each requirement gets an upload slot. Priya drops her passport, and invitation letter (for demo purposes, all other files are auto-loaded after the first 2 files manually loaded).

The Document Intelligence Agent reads each one using vision, in any language, and cross-checks them against the requirements in real time.

After the AI reads in and analyzes each travel document, the advisory agent then begins the processing of providing a plan to fix errors in the traveler's initial application submission.

The advisory agent presents fixes for complex cross-lingual mistake, ex: Priya was invited to Munich by a European conference committee, but the German consulate need an official business invitation from a named German company on their letterhead. Each fix comes with specific, actionable instructions — not just what's wrong, but exactly what to do.

Each demo profile incorporates suggested edits to be made to existing travel documents. Priya clicks Apply Fixes & Re-check. The re-audit pipeline verifies each corrected document through four steps (Fetching → Reading → Cross-checking → Verifying). All three fixes pass.

This works for any corridor. Here's a completely different one — Azerbaijan to Guinea-Bissau — with the Research Agent live-searching government sources it's never seen before.

Architecture

Traverse runs three AI agents on Claude Opus 4.6, coordinated by a TypeScript orchestrator that streams results to the UI in real-time via Server-Sent Events.

graph LR
    INPUT["User Input<br/>corridor + documents"]

    subgraph ORCH["Orchestrator"]
        direction TB
        PLAN["Plan agents"]
        PARALLEL["Parallel dispatch"]
        WAIT["Wait for both"]
        SEQ["Sequential pass"]
    end

    subgraph R["Research Agent"]
        direction TB
        R1["Web Search<br/><i>live gov sources</i>"]
        R2["Extended Thinking<br/><i>reason about policies</i>"]
        R3["RequirementsChecklist<br/><i>5-12 items + metadata</i>"]
        R1 --> R2 --> R3
    end

    subgraph D["Document Intelligence Agent"]
        direction TB
        D1["Pass 1: Vision Read<br/><i>OCR every document</i>"]
        D2["Pass 2: Cross-Analysis<br/><i>compliance + contradictions</i>"]
        D3["DocumentAnalysis<br/><i>compliance, cross-lingual,<br/>narrative, forensics</i>"]
        D1 --> D2 --> D3
    end

    subgraph A["Advisory Agent"]
        direction TB
        A1["Extended Thinking<br/><i>16K token budget</i>"]
        A2["Warm Guidance<br/><i>actionable fixes + tips</i>"]
        A3["AdvisoryReport<br/><i>fixes, tips, warnings</i>"]
        A1 --> A2 --> A3
    end

    INPUT --> PLAN
    PLAN --> PARALLEL
    PARALLEL -->|"parallel"| R1
    PARALLEL -->|"parallel"| D1
    R3 --> WAIT
    D1 --> WAIT
    WAIT --> SEQ
    SEQ -->|"requirements<br/>+ extractions"| D2
    D3 -->|"compliance data"| A1

    style R fill:#eff6ff,stroke:#3b82f6,stroke-width:2px
    style D fill:#f5f3ff,stroke:#8b5cf6,stroke-width:2px
    style A fill:#ecfdf5,stroke:#10b981,stroke-width:2px
    style ORCH fill:#fff7ed,stroke:#f59e0b,stroke-width:2px

All three agents use Claude Opus 4.6 with extended thinking — dynamically allocating reasoning depth based on complexity.

Agent	Opus 4.6 Capabilities	What It Does
Research	Web Search, Extended Thinking	Searches live government sources for corridor-specific visa requirements
Document Intelligence	Vision, Extended Thinking	Reads documents in any language, cross-references for contradictions
Advisory	Extended Thinking, Structured Output	Synthesizes findings into prioritized, actionable fixes

Demo Walkthrough — Priya Sharma

The demo follows Priya, a software engineer in Bangalore applying for a business visa to Germany. Her nine documents contain a cross-lingual contradiction she doesn't know about.

graph TD
    subgraph SETUP["Setup"]
        S1["Open Sidebar"]
        S2["Select Priya Sharma"]
        S3["Load Profile<br/><i>India → Germany, Business</i>"]
        S4["Form Auto-fills<br/><i>animated field slide</i>"]
        S5["Click Check Requirements"]
        S1 --> S2 --> S3 --> S4 --> S5
    end

    subgraph RESEARCH["Phase 1: Research"]
        R1["Research Agent Activates<br/><i>cached india-germany.json</i>"]
        R2["Sources Found<br/><i>German Embassy, VFS Global</i>"]
        R3["12 Requirements Loaded<br/><i>9 uploadable, 3 info-only</i>"]
        R4["Phase 1 Advisory<br/><i>instant — no LLM</i>"]
        R1 --> R2 --> R3 --> R4
    end

    subgraph DOCS["Phase 2: Document Verification"]
        D1["Passport Uploaded"]
        D2["Auto-Cascade<br/><i>8 remaining docs, 300ms stagger</i>"]
        D3["9 Parallel Vision Reads<br/><i>Opus 4.6 + Vision</i>"]
        D4["9 Cross-Checks<br/><i>compliance per requirement</i>"]
        D1 --> D2 --> D3 --> D4
    end

    subgraph FINDINGS["Findings"]
        F1["CRITICAL<br/>Cross-lingual contradiction<br/><i>Hindi: permanent ₹85K<br/>English: contract ₹60K</i>"]
        F2["CRITICAL<br/>Gmail forensic flag<br/><i>easummit2026@gmail.com</i>"]
        F3["WARNING<br/>Narrative gap<br/><i>15-day trip, 3-day conference</i>"]
    end

    subgraph ADVISORY["Phase 3: Advisory"]
        A1["Phase 1b Updates<br/><i>advisory evolves in real-time</i>"]
        A2["Phase 2 LLM Synthesis<br/><i>Opus 4.6 + Extended Thinking</i>"]
        A3["Modal Opens<br/><i>Let's Strengthen Your Application</i>"]
        A1 --> A2 --> A3
    end

    subgraph REAUDIT["Phase 4: Re-Audit"]
        X1["Apply Fixes & Re-check"]
        X2["Fix 1: Cover Letter<br/><i>Corrected to match Hindi docs</i><br/>PASS"]
        X3["Fix 2: Invitation<br/><i>Official domain + letterhead</i><br/>PASS"]
        X4["Fix 3: Itinerary<br/><i>Day-by-day plan added</i><br/>PASS"]
        X5["Looking Good!<br/><i>APPLICATION PROCEEDS</i>"]
        X6["Switch to Hindi<br/><i>नमस्ते — modal translates</i>"]
        X1 --> X2 --> X3 --> X4 --> X5 --> X6
    end

    S5 --> R1
    R4 --> D1
    D4 --> F1
    D4 --> F2
    D4 --> F3
    D4 --> A1
    F1 & F2 & F3 --> A2
    A3 --> X1

    style SETUP fill:#f8fafc,stroke:#64748b,stroke-width:1px
    style RESEARCH fill:#eff6ff,stroke:#3b82f6,stroke-width:2px
    style DOCS fill:#f5f3ff,stroke:#8b5cf6,stroke-width:2px
    style FINDINGS fill:#fef2f2,stroke:#ef4444,stroke-width:2px
    style ADVISORY fill:#ecfdf5,stroke:#10b981,stroke-width:2px
    style REAUDIT fill:#f0fdf4,stroke:#16a34a,stroke-width:2px

    style F1 fill:#fee2e2,stroke:#dc2626
    style F2 fill:#fee2e2,stroke:#dc2626
    style F3 fill:#fef3c7,stroke:#d97706
    style X5 fill:#bbf7d0,stroke:#16a34a,stroke-width:2px

SSE Event Flow (Technical Detail)

sequenceDiagram
    participant C as Client (Browser)
    participant S as /api/analyze
    participant R as Research Agent
    participant D as Document Agent
    participant A as Advisory Agent

    C->>S: POST /api/analyze {travelDetails, documents}
    activate S

    S->>R: runResearchAgent()
    S->>D: runDocumentReaderPass1()
    Note over R,D: Parallel execution

    R-->>S: orchestrator/agent_start
    S-->>C: SSE: orchestrator/agent_start [Research]

    R-->>S: search_status (searching)
    S-->>C: SSE: search_status
    R-->>S: search_status (found)
    S-->>C: SSE: search_status

    R-->>S: thinking (excerpt)
    S-->>C: SSE: thinking [Research]

    loop For each requirement
        R-->>S: requirement
        S-->>C: SSE: requirement
    end

    R-->>S: orchestrator/agent_complete
    S-->>C: SSE: orchestrator/agent_complete [Research]

    D-->>S: document_read (doc 1)
    S-->>C: SSE: document_read
    D-->>S: document_read (doc N)
    S-->>C: SSE: document_read
    D-->>S: orchestrator/agent_complete
    S-->>C: SSE: orchestrator/agent_complete [Document Reader]

    Note over S: Both complete — start Pass 2

    S->>D: runDocumentAnalyzerPass2()
    D-->>S: thinking (analysis)
    S-->>C: SSE: thinking [Document Analysis]
    D-->>S: cross_lingual
    S-->>C: SSE: cross_lingual
    D-->>S: forensic
    S-->>C: SSE: forensic
    D-->>S: narrative
    S-->>C: SSE: narrative

    S-->>C: SSE: complete {requirements, extractions, analysis}
    deactivate S

    Note over C: Phase 2 triggers automatically

    C->>A: POST /api/advisory
    activate A
    A-->>C: SSE: recommendation (fix 1)
    A-->>C: SSE: recommendation (fix 2)
    A-->>C: SSE: advisory_tips
    A-->>C: SSE: assessment
    A-->>C: SSE: orchestrator/agent_complete [Advisory]
    deactivate A

    Note over C: Advisory modal auto-opens

Key Features

🌍 Universal Coverage — Works for any travel corridor (37,830+ combinations), any purpose (tourism, business, study, work, medical, family)
🗣️ 40 Languages — Upload documents in any language; get results translated to your preferred language
🔍 Three-Phase Analysis — Research → Document Intelligence → Advisory synthesis
⚡ Minimal Dead Time — Two-phase advisory shows instant preliminary results, updates progressively, then refines with AI
🔒 Privacy-First — No database, no accounts, documents processed in-memory and immediately discarded
🎨 Beautiful UX — Dual themes, progressive translation, real-time AI reasoning display, smooth animations

How It Works

Phase 1: Research Agent (Opus 4.6 + Web Search + Extended Thinking)

Searches live government sources for current visa requirements
Uses extended thinking to reason about complex visa policies
Returns structured checklist with requirement details, sources, fees, processing times
Falls back to cached data for instant results if live search unavailable

Phase 2: Document Intelligence Agent (Opus 4.6 + Vision + Extended Thinking)

Pass 1: Reads all documents in parallel using multimodal vision
- Extracts text from images in any language
- Identifies document types (passport, bank statement, invitation letter, etc.)
- Detects language and structured data
Pass 2: Cross-document analysis with extended thinking
- Checks compliance against requirements
- Detects cross-lingual contradictions (e.g., employment status differs between languages)
- Assesses narrative coherence (dates, amounts, story consistency)
- Flags potential document forensics issues

Phase 3: Advisory Agent (Two-Phase Architecture)

Phase 3a (Instant): Programmatic advisory built from requirements
- Shows immediately while documents are analyzing
- Lists all required documents and initial guidance
Phase 3b (Progressive): Updates as each document completes
- Severity changes based on compliance (info → warning → critical)
- Overall assessment evolves in real-time
Phase 3c (Refined): Opus 4.6 synthesis with extended thinking
- Triggers after 80% of documents analyzed
- Refines recommendations with specific, actionable guidance
- Adds helpful URLs (embassy sites, government portals)
- Generates personalized interview tips
- Warm, encouraging tone: "Let's fix this" not "You must correct"

Tech Stack

Frontend

Next.js 16.1.6 (App Router)
React 19.2.3
TypeScript 5
Tailwind CSS 4 + shadcn/ui components
Server-Sent Events (SSE) for real-time streaming

Backend

Anthropic SDK 0.74.0 (claude-opus-4-6)
- Extended thinking (16K token budget)
- Web search (Research Agent)
- Vision (Document Intelligence Agent)
- Streaming responses
Node.js runtime
Zod for runtime validation
In-memory rate limiting
Environment variable validation

Infrastructure

Vercel deployment
No database — stateless, privacy-first architecture
Edge-compatible API routes

Demo Personas

Try one of three pre-built demo personas with realistic document sets:

Persona	Corridor	Purpose	Documents	Key Issues
Priya Sharma	India → Germany	Business (tech conference)	9 docs (Hindi + English)	Cross-lingual employment contradiction, low savings ratio
Amara Okafor	Nigeria → UK	Study (MSc)	8 docs (English)	Standard student visa requirements
Carlos Mendes	Brazil → Japan	Tourism (photography)	6 docs (Portuguese + English)	Freelance income documentation

Click the personas in the sidebar to auto-load their profiles and documents.

Getting Started

Prerequisites

Node.js 18+ and npm
Anthropic API key (get one here)

Installation

# Clone the repository
git clone <repository-url>
cd clearpath

# Install dependencies
npm install

# Set up environment variables
# Create a new .env.local file
touch .env.local

Edit .env.local and add your Anthropic API key:

ANTHROPIC_API_KEY=sk-ant-api03-...

Running Locally

npm run dev

Open http://localhost:3000 in your browser.

Building for Production

npm run build
npm start

Environment Variables

Variable	Required	Default	Description
`ANTHROPIC_API_KEY`	Yes	-	Your Anthropic API key
`ANTHROPIC_MODEL`	No	`claude-opus-4-6`	Model for Research & Document agents
`ADVISORY_MODEL`	No	`claude-opus-4-6`	Model for Advisory agent
`USE_LIVE_SEARCH`	No	`false`	Enable live web search (otherwise uses cached data)
`NODE_ENV`	No	`development`	Environment (development/production)

Security & Production Features

Security

Environment Validation — Runtime validation of required environment variables with clear error messages
API Key Protection — Never committed to version control, validated on startup
Rate Limiting — In-memory rate limiter protects all API endpoints:
- /api/analyze: 10 requests/minute (STRICT)
- /api/translate: 30 requests/minute (STANDARD)
- /api/advisory: 30 requests/minute (STANDARD)
Type Safety — Full TypeScript with Zod runtime validation for all inputs
Error Handling — Comprehensive error boundaries and graceful degradation

Development

Development-Only Logging — Console statements gated with isDevelopment() checks
No Console Pollution in Production — Clean production builds
Type Guards — Proper type guards instead of as any assertions
Named Constants — All magic numbers extracted to configuration

Privacy

No database. No user data stored anywhere.
No accounts. No login, no signup, no tracking cookies.
In-memory only. Documents are base64-encoded in the browser, sent to the API route, processed, and immediately discarded.
No server-side file storage. All processing happens in memory during the request.
No tracking. No analytics, no third-party scripts, no cookies.

Demo Documents

The demo-docs/ directory contains synthetic documents for demonstration purposes:

26 realistic document templates (HTML)
Rendered to PNG using Puppeteer
All data is completely fictional — no real PII anywhere
Documents in multiple languages (Hindi, Portuguese, English)
Intentional flaws for demonstration (contradictions, low amounts, etc.)

Project Structure

clearpath/
├── src/
│   ├── app/                      # Next.js App Router
│   │   ├── page.tsx             # Landing page with travel form
│   │   ├── analyze/page.tsx     # Analysis page with live results
│   │   └── api/                 # API routes
│   │       ├── analyze/route.ts # Main orchestrator (SSE)
│   │       ├── advisory/route.ts # Phase 2 advisory (SSE)
│   │       └── translate/route.ts # Translation endpoint
│   ├── lib/                      # Core logic
│   │   ├── agents/              # AI agent implementations
│   │   │   ├── research.ts      # Research Agent (Opus 4.6 + Web Search)
│   │   │   ├── document.ts      # Document Intelligence Agent (Opus 4.6 + Vision)
│   │   │   └── advisory.ts      # Advisory Agent (Opus 4.6 + Thinking)
│   │   ├── advisory-builder.ts  # Phase 1 & 1b programmatic advisory
│   │   ├── env.ts               # Environment validation
│   │   ├── rate-limit.ts        # Rate limiting middleware
│   │   ├── validation.ts        # Zod schemas
│   │   ├── types.ts             # TypeScript types
│   │   ├── config.ts            # Configuration constants
│   │   ├── i18n-context.tsx     # Translation system
│   │   └── demo-context.tsx     # Demo persona state
│   └── components/              # React components (19 files)
│       ├── advisory-modal.tsx   # Refined advisory modal
│       ├── advisory-card.tsx    # Progressive advisory card
│       ├── live-feed.tsx        # SSE event renderer
│       ├── thinking-panel.tsx   # Real AI reasoning display
│       └── ...
├── demo-docs/                   # Synthetic demo documents
│   ├── templates/               # HTML templates
│   └── personas/                # Rendered PNGs
└── data/                        # Cached fallback data
    └── corridors/               # Pre-researched visa requirements

Development

Key Scripts

npm run dev          # Start development server
npm run build        # Build for production
npm run start        # Run production build
npm run lint         # Run ESLint

Built With Opus 4.6

This project showcases Claude Opus 4.6 capabilities:

Extended Thinking (16K budget) — All three agents reason through complex visa policies
Web Search — Research Agent searches live government sources
Vision — Document Intelligence Agent reads documents in any language
Multi-Agent Coordination — TypeScript orchestrator manages three specialized agents
Streaming — Real-time SSE streams show thinking and results as they happen

Built for an Anthropic hackathon.

Troubleshooting

"ANTHROPIC_API_KEY is not set"

Make sure you created .env.local (not .env)
Check that the API key starts with sk-ant-
Restart the dev server after adding the key

"Rate limit exceeded"

Development mode bypasses rate limits
Production limits: 10 req/min for analysis, 30 req/min for translation
Wait for the retry-after period shown in the error

Documents not uploading

Max file size: 5MB per file
Supported formats: PNG, JPEG only
Max 12 files total
Check browser console for errors

Translation not working

Ensure language selector shows current language
Translation happens in 3 phases (UI → Content → Live Feed)
Check network tab for /api/translate requests
English is default, no translation needed

Extended thinking not showing

Thinking panels appear during analysis (collapsible)
Look for the eye icon "View AI reasoning" on each requirement
Thinking text shows in real-time during processing
Check that ANTHROPIC_MODEL=claude-opus-4-6 in .env.local

Acknowledgments

Built with Claude Opus 4.6 by Anthropic
UI components from shadcn/ui
Icons from Lucide
Inspired by the millions of visa applicants worldwide who face preventable rejections

License

MIT

Name		Name	Last commit message	Last commit date
data/corridors		data/corridors
demo-docs		demo-docs
public		public
screen-shots		screen-shots
src		src
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
components.json		components.json
eslint.config.mjs		eslint.config.mjs
next.config.ts		next.config.ts
package-lock.json		package-lock.json
package.json		package.json
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json
vercel.json		vercel.json

Latest commit

History

Folders and files

Repository files navigation

Traverse

Quick Start

How It Looks

Architecture

Demo Walkthrough — Priya Sharma

Key Features

How It Works

Phase 1: Research Agent (Opus 4.6 + Web Search + Extended Thinking)

Phase 2: Document Intelligence Agent (Opus 4.6 + Vision + Extended Thinking)

Phase 3: Advisory Agent (Two-Phase Architecture)

Tech Stack

Frontend

Backend

Infrastructure

Demo Personas

Getting Started

Prerequisites

Installation

Running Locally

Building for Production

Environment Variables

Security & Production Features

Security

Development

Privacy

Demo Documents

Project Structure

Development

Key Scripts

Built With Opus 4.6

Troubleshooting

"ANTHROPIC_API_KEY is not set"

"Rate limit exceeded"

Documents not uploading

Translation not working

Extended thinking not showing

Acknowledgments

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages