PatrickAttankurugu/FinCrime-LLM

AI-powered financial crime detection for African markets using fine-tuned Mistral 7B # FinCrime-LLM **A Large Language Model (LLM) Project For Financial Crime Detection in African Markets** This project aims to fine-tune the Mistral 7B model to generate Suspicious Activity Reports (SARs), KYC assessments, and transaction analysis tailored to African financial contexts. Currently in active development. ## Project Status **Work in Progress**: This model is currently under development. Training and deployment have not yet been completed. The codebase provides the infrastructure and workflows needed for: - Synthetic data generation - Model fine-tuning with QLoRA - API and demo interfaces (ready for when the model is trained) - Testing and evaluation frameworks ## Planned Features - **SAR Generation**: Will automatically generate comprehensive Suspicious Activity Reports - **KYC Assessment**: Will perform risk-based customer due diligence - **Transaction Analysis**: Will detect suspicious patterns and red flags - **QLoRA Training**: Efficient 4-bit quantized training on consumer GPUs - **WandB Integration**: Track experiments and model performance - **Compliance-First**: Building with regulatory requirements in mind ## Planned Architecture ```mermaid graph TB A[Raw Data] -->|Synthetic Generation| B[OpenAI GPT-4] B -->|JSON SARs| C[Data Processing] C -->|Instruction Format| D[Training Dataset] D -->|QLoRA Fine-tuning| E[Mistral 7B] E -->|Model to be Trained| F[Future Production Model] F --> G[FastAPI Backend] F --> H[Streamlit Demo] G --> I[SAR Generator] G --> J[KYC Assessor] G --> K[Transaction Analyzer] G --> L[Compliance Checker] M[User] --> H N[Client Apps] --> G style F fill:#FFA500 style G fill:#2196F3 style H fill:#FF9800 ``` ## Project Structure ``` FinCrime-LLM/ ├── api/ # FastAPI backend │ ├── routers/ # API endpoints (SAR, KYC, Transaction, Compliance) │ ├── models/ # Pydantic schemas │ └── utils/ # Auth, logging utilities ├── data/ │ ├── raw/ # Raw financia …