Neural Financial Intelligence & Private Ledger Engine
Cognis is a professional-grade financial intelligence engine designed to replace cloud-dependent budgeting apps with a locally hosted, AI-driven architecture. By processing raw financial data through local Large Language Models (LLMs), Cognis provides deep insights into spending patterns without compromising sensitive personal data.
In an era of data harvesting, Cognis was built to prove that sophisticated financial analysis doesn't require cloud exposure. The application handles the end-to-end lifecycle of financial data: from raw CSV ingestion and AI-augmented semantic labeling to persistent storage and real-time visualization.
The core philosophy is Local-First Intelligence. By utilizing a local Ollama instance, Cognis performs complex natural language processing to categorize messy bank descriptions into clean, actionable data points, ensuring that your financial "footprint" never leaves your physical hardware.
Cognis utilizes a decoupled client-server architecture designed for low latency and high data integrity.
- Semantic Analysis: Raw transaction strings are fed into a localized Llama 3 model.
- Contextual Inference: The AI interprets ambiguous vendor names (e.g., "SQ *PLUMBING") and maps them to logical categories (e.g., "Home Maintenance").
- Stateless Processing: API endpoints handle file uploads and data retrieval via asynchronous I/O.
- Data Normalization: Uses Pandas to perform de-duplication, ensuring that redundant uploads do not corrupt the financial ledger.
- Immutable Persistence: Implements absolute pathing logic on the host filesystem to maintain a consistent data state across server restarts and environment shifts.
- State Synchronization: Utilizes React hooks (
useCallback,useEffect) to ensure the UI stays in sync with the backend CSV state. - Reactive Visualization: Implements Recharts for dynamic data modeling and CSS-based "Dark Mode" optimized for professional workstations.
| Layer | Technology |
|---|---|
| Intelligence | Ollama Core (Llama 3 / Phi-3) |
| Data Engine | Python 3.10+, Pandas, Absolute FS |
| Interface | React 18, Tailwind CSS, Recharts |
| API Layer | FastAPI (Asynchronous I/O) |
- Hardware: Optimized for local LLM execution (8GB+ RAM recommended).
- Environment: Python 3.10+, Node.js 18+.
- LLM Host: Ollama must be installed and running
llama3.
cd backend
python -m venv venv
# Windows
venv\Scripts\activate
# Linux/Mac
source venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --port 8000 --reload