Arthyx: Autonomous Quantitative Financial Analyst

Arthyx is an advanced, industry-grade autonomous agent tailored for quantitative financial analysis. Its name is derived from the Sanskrit word "Arth", signifying Meaning, Wealth, and Finance, reflecting its core mission: to derive actionable financial intelligence from complex unstructured data.

Unlike standard retrieval-augmented generation (RAG) systems that function as simple document summarizers, Arthyx is engineered as a comprehensive analytical platform. It integrates semantic search, probabilistic reasoning, and rigid regulatory frameworks to perform work equivalent to a human quantitative analyst. It is designed to rigorously handle the complexity of institutional finance, including Basel III norms, RBI regulations, and credit risk modeling.

Unique Value Proposition

Arthyx stands apart through its hybrid architecture that combines the creative reasoning of Large Language Models with the deterministic accuracy of Knowledge Graphs and Quantitative Models.

Deep Contextual Understanding

Most RAG systems retrieve small, fragmented snippets of text (200-500 characters), leading to hallucinated or incomplete answers. Arthyx employs a "Deep Context" engine that retrieves massive 3000-character chunks and aggregates the top-15 most relevant sections. This allows it to "read" approximately 15-20 full pages of text before generating an answer, ensuring that every response is synthesized from a comprehensive understanding of the document rather than isolated keywords.

High-Fidelity Visual Analysis

Arthyx automatically detects numerical trends in financial data and generates precise, interactive visualizations. Whether it is a 5-year comparison of Non-Performing Assets (NPA) or a breakdown of Capital Adequacy Ratios (CRAR), the system visualizes the data instantly. These charts are dynamic and state-aware; editing a previous query automatically recalculates the underlying data and redraws the visualization to reflect the new context, ensuring perfect synchronization between conversation history and visual output.

Knowledge Graph Auditing

Beyond simple text search, Arthyx maps entities (companies, directors, subsidiaries) into a Neo4j Knowledge Graph. This allows it to uncover hidden relationships, conflicts of interest, and risk contagion paths that purely vector-based systems would miss. It effectively performs a "Knowledge Audit" on every uploaded document.

Quantitative Risk Modeling

The system includes a dedicated risk engine that extracts financial ratios from unstructured text and evaluates them against encoded regulatory standards (Basel III, RBI Master Directions). It calculates scores for Credit Risk, Market Risk, and Operational Risk, providing a structured quantitative assessment alongside the qualitative text analysis.

Specialized Indian Language OCR

Recognizing the diverse landscape of Indian finance, Arthyx features a specialized OCR pipeline fine-tuned for Indian languages. It achieves high accuracy on Hindi, Tamil, Bengali, Gujrati, and Telugu scripts, making it the only open-source financial agent capable of auditing regional vernacular records with the same precision as English documents.

Technical Architecture

The system is built on a production-ready stack designed for scale, speed, and type safety.

Core Intelligence

• Google Gemini 2.0 Flash: Selected for its massive context window and reasoning capabilities.
• text-embedding-004: High-dimensional vectorization for semantic nuance.

Data Infrastructure

• Pinecone Serverless: Low-latency vector retrieval (<50ms).
• Neo4j AuraDB: Graph database for entity relationship modeling.
• Upstash Redis: High-performance caching for session state and async job queues.

Application Layer

• Next.js 15 (Turbopack): React Server Components and Edge handling.
• Async Polling Architecture: A robust job-queue system that handles massive 500+ page PDFs without HTTP timeouts, ensuring 100% reliability for large annual reports.

Installation and Deployment

Arthyx is open-source and ready for deployment on any Node.js compatible environment.

Prerequisites

• Node.js 18+
• Accounts for Google AI, Pinecone, Upstash, and Neo4j.

Local Development

1. Clone the Repository

   git clone https://github.com/IndAlok/Arthyx.git
   cd Arthyx

2. Install Dependencies

   npm install --legacy-peer-deps

3. Configure Environment Create a .env.local file with your credentials:

   GOOGLE_API_KEY=your_key
   PINECONE_API_KEY=your_key
   PINECONE_INDEX_HOST=your_index_host
   NEXT_PUBLIC_SUPABASE_URL=https://<project-ref>.supabase.co
   NEXT_PUBLIC_SUPABASE_ANON_KEY=your_publishable_anon_key
   NEO4J_URI=your_uri
   NEO4J_USERNAME=your_username
   NEO4J_PASSWORD=your_password
   UPSTASH_REDIS_REST_URL=your_url
   UPSTASH_REDIS_REST_TOKEN=your_token

4. Launch Application

   npm run dev

Cloudflare Pages Deployment (recommended)

This project is configured for Cloudflare Pages using @cloudflare/next-on-pages.

Cloudflare Pages build settings

• Framework preset: Next.js
• Build command: npm run build
• Build output directory: .vercel/output/static
• Compatibility flags: enable nodejs_compat (Preview + Production)

Deployment notes:

• Cloudflare Pages deploys automatically after the build finishes. You do not need a deploy command.
• If you previously set a Pages “Deploy command” like npm run deploy, this repo now makes that command a safe no-op on Pages builds (so token/auth errors can’t break your deployment).

Environment variables

Set these in Cloudflare Pages → Settings → Environment variables (Preview + Production):

• GOOGLE_API_KEY
• PINECONE_API_KEY
• PINECONE_INDEX_HOST (your Pinecone index host, e.g. my-index-xxxx.svc.<region>.pinecone.io)
• UPSTASH_REDIS_REST_URL
• UPSTASH_REDIS_REST_TOKEN
• NEXT_PUBLIC_SUPABASE_URL
• NEXT_PUBLIC_SUPABASE_ANON_KEY

Optional (recommended if you want to automatically delete uploaded files from Supabase Storage after indexing to save storage):

• SUPABASE_SERVICE_ROLE_KEY (server-only; Supabase dashboard “Secret key”).

Notes:

• Never put SUPABASE_SERVICE_ROLE_KEY in any NEXT_PUBLIC_* variable.

Notes:

• Because Cloudflare Pages runs API routes on the Edge runtime, Word/Excel extraction is handled via Gemini (not Node-based parsers).

Optional (only if you actually want Neo4j features):

• NEO4J_URI
• NEO4J_USERNAME
• NEO4J_PASSWORD

Local Cloudflare preview

@cloudflare/next-on-pages is known to be unreliable on Windows. For local preview, run under WSL/Linux/macOS:

npm run build
npm run preview

Optional: manual/CI deploy with Wrangler

If you want to deploy from your own machine or CI (outside Cloudflare Pages), you can run:

npm run build
RUN_WRANGLER_DEPLOY=1 npm run deploy

This requires Wrangler authentication (either wrangler login or a CLOUDFLARE_API_TOKEN with permission to deploy to Cloudflare Pages projects).

On Windows PowerShell:

$env:RUN_WRANGLER_DEPLOY='1'
npm run deploy

Contributing

We welcome contributions from engineers passionate about quantitative finance and AI. To contribute:

1. Fork the repository.
2. Create a feature branch for your specific improvement.
3. Commit your changes with clear, semantic messages.
4. Submit a Pull Request detailing the technical implementation.

---

Engineered by IndAlok | Algorithmic Financial Intelligence for the Specialist.