---

📍 Problem Statement

Traditional credit underwriting is plagued by fragmented data, slow manual research, and "black-box" decisioning. Banks lose precious time manually parsing complex financial tables and searching for litigation records.

💡 Our Solution

An autonomous AI agent that handles the end-to-end credit appraisal process. It doesn't just "process" data—it reasons through it, conducts live web research, verifies facts via Human-in-the-Loop (HITL), and generates a professional Credit Appraisal Memo (CAM) backed by a high-accuracy ML model.

---

✨ Key Pillars of Intelligence

• 🧠 Dual-Brain Architecture: Separates Orchestration (Llama 3.3/Gemini 3 pro) from Deep Analysis (Gemini 2.5 and 3 Pro).
• 🧬 High-Fidelity Extraction: LlamaParse + Pinecone Cloud convert complex PDFs into searchable Markdown, reducing hallucinations by 85%.
• 🌐 Granular Web Research: Autonomous one-by-one scrutiny of search results via Tavily to map NCLT filings and RBI penalties.
• 🤖 Predictive Decisioning: Pre-trained XGBoost Classifier (97% accuracy) predicts Approval, Limits, and Interest Rates.
• 🤝 Human-in-the-Loop (HITL): Sequential Review Panels ensure human oversight and data correction at every critical step.

---

⚡ The 5-Phase Systematic Workflow

1. Phase 1: Sequential Document Intelligence – Automated verification of mandatory docs + 5Cs Insights extraction.
2. Phase 2: External Risk Discovery – Granular litigation search and live news sentiment cross-verification.
3. Phase 3: Numerical Feature Engineering – Locking extracted metrics into sync with the main Vault; supports manual overrides.
4. Phase 4: ML Scoring & Decisioning – Running the XGBoost engine to determine creditworthiness and limits.
5. Phase 5: Automated CAM Generation – Drafting and exporting a high-fidelity PDF Credit Appraisal Memorandum.

---

🏗️ System Workflow

graph TD
    Start["    🚀 User Uploads Docs & Model Select    "] --> P1["    📂 Phase 1: RAG Analysis    "]
    P1 <--> H1["&nbsp;&nbsp;&nbsp; 💬 HITL Review &nbsp;&nbsp;&nbsp;"]
    P1 --> P2["&nbsp;&nbsp;&nbsp; 🌐 Phase 2: Web Litigation &nbsp;&nbsp;&nbsp;"]
    P2 <--> H2["&nbsp;&nbsp;&nbsp; 💬 HITL Review &nbsp;&nbsp;&nbsp;"]
    P2 --> P3["&nbsp;&nbsp;&nbsp; 🔢 Phase 3: Feature Extraction &nbsp;&nbsp;&nbsp;"]
    P3 <--> H3["&nbsp;&nbsp;&nbsp; 💬 HITL Review &nbsp;&nbsp;&nbsp;"]
    P3 --> P4["&nbsp;&nbsp;&nbsp; ⚖️ Phase 4: ML Scoring &nbsp;&nbsp;&nbsp;"]
    P4 <--> H4["&nbsp;&nbsp;&nbsp; 💬 HITL Review &nbsp;&nbsp;&nbsp;"]
    P4 --> P5["&nbsp;&nbsp;&nbsp; 📄 Phase 5: CAM Export &nbsp;&nbsp;&nbsp;"]
    P5 --> Done["&nbsp;&nbsp;&nbsp; ✅ Final Confirmation &nbsp;&nbsp;&nbsp;"]

Loading

---

🛠️ Tech Stack

Layer	Technology
Orchestration	LangGraph (Stateful ReAct Pattern)
Reasoning	Llama 3.3 (Groq) / Gemini 2.5 Flash
RAG / Memory	Pinecone Cloud + LlamaParse
Web Intel	Tavily AI (Credit Research Mode)
ML Engine	XGBoost (Binary Classification + Regression)
Interface	Streamlit (Stateful Chat & UI)

---

🚀 Getting Started

# Clone & Setup
git clone https://github.com/ShivamMaurya14/CREDI-MITRA.git
cd CREDI-MITRA
pip install -r requirements.txt

# Configure .env
PINECONE_API_KEY=pcsk_...
GROQ_API_KEY=gsk_...
GOOGLE_API_KEY=AIza...
TAVILY_API_KEY=tvly-...
LLAMA_CLOUD_API_KEY=llx-...

# Launch
streamlit run app.py

---

🔮 Roadmap

• [*] Multi-Model Support: Selection of Orchestrator/Analyst.
• [*] Pinecone RAG: Long-term vector memory for financial data.
• Email Integration: Automated Acceptance/Rejection alerts via SendGrid.
• Database Autofetch: Ingest historical records via Application No.
• API Pulls: Real-time GST/MCA/CIBIL verification via official APIs.
• LiveKit Voice: AI-driven borrower interviews (Character 5C).
• Blockchain Ledger: Immutable logs of all AI credit decisions.

Developed for Hackathon 2026 Built with ❤️ by Shivam Maurya