Credit Decision Memory Engine (Prototype)

An explainable, similarity-driven credit decisioning system that uses historical "memory" to assess risk, detect anomalies, and provide evidence-based recommendations.

🧠 Memory-Based Decisioning?

Traditional systems use rigid rules or opaque black-box models. This engine works like a human underwriter with a perfect memory:

1. Recalls the most similar past applications.
2. Compares the current outcome to those historical precedents.
3. Explains the decision by citing the specific past cases ("Approved because it looks like Case #124 and #98 which repaid successfully").
4. Detects Anomalies if an application looks nothing like what we've seen before.

🚀 Key Features

• Multi-Factor Profiling: Combines income, expenses, loan details, and credit history into a unified vector.
• Top-K Retrieval: instantly finds the nearest historical analogs using cosine/euclidean distance.
• Anomaly Detection: Flags applications that are statistical outliers or match known fraud patterns.
• Audit Trail: Every decision returns the "Evidence" (IDs of cases used to make the decision).
• Learning Loop: The system gets smarter as more decisions are finalized and fed back into memory.

🏗️ Architecture

This project uses Qdrant as its core memory store.

• Vectors: All applications are converted into high-dimensional vectors.
• Search: We use Qdrant's similarity search to find the "Nearest Neighbors" in the vector space.
• Decision: The decision is a weighted vote of these neighbors.

📂 Project Structure

src/
├── similarity.py       # Vectorization logic (Income -> normalized vector)
├── memory.py           # In-memory vector store & Top-K retrieval
├── anomaly.py          # Outlier detection & 'Fraud' pattern matching
└── decision_engine.py  # Orchestrator: Profile -> Vector -> Memory -> Decision

data/
└── history.json        # Synthetic historical dataset (generated)

demo/
└── run_demo.py         # End-to-end CLI demonstration script

🛠️ Setup & Usage

Prerequisites

• Python 3.8+
• numpy, scikit-learn

Installation

pip install -r requirements.txt

Running the Demo

1. Generate Data (Creates synthetic history)

   python data/history_generator.py

2. Run the Chatbot (Interactive)

   python src/chatbot.py

   Interact naturally: "Assess income=15000 credit=good"

📊 Example Output

Application: {'income': 12000, 'credit_history': 'good', ...}

>>> Recommendation: APPROVE (Confidence: 1.0)
Explanation: Based on 5 similar past cases (5 approved). Most similar case was approve.
Similar Cases:
  - Case 42 (approve) [Dist: 0.12]: Similar Income, Same Credit Profile
  - Case 15 (approve) [Dist: 0.15]: Similar Income, Same Credit Profile

🔮 Future Improvements

• Embeddings Model: Replace manual vectorization with a trained Neural Network (e.g., TabTransformer) or LLM embeddings for documents.
• Vector Database: Migrate from in-memory list to Milvus/Pinecone for million-scale retrieval.
• Document Analysis: Integrate OCR to actually parse uploaded bank statements instead of stubbing them.