Skip to content

This project, Auditron, was created as part of the Challenge-Based Learning (CBL) AI project at Esprit School of Engineering - Tunisia.

Notifications You must be signed in to change notification settings

fouratmansouri/Auditron

Folders and files

NameName
Last commit message
Last commit date

Latest commit

ย 

History

47 Commits
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 

Repository files navigation

๐Ÿค– Auditron README ๐Ÿงฎ

Auditron Logo

๐Ÿ“‹ Overview

This project, Auditron, aims to develop an AI-powered financial audit assistant (chatbot) to streamline compliance verification and tax calculations for businesses. It is part of a semester-long academic project at Esprit School of Engineering, Tunisia, which tackles Agetnic AI. This tool will combine legal expertise (constitutional and local finance laws) with mathematical precision to automate critical audit tasks, reducing human error and operational costs.

โœจ Features

๐Ÿ› ๏ธ Tech Stack

๐Ÿ’ป Frontend

CSS + HTML

โš™๏ธ Backend

Flask

๐Ÿงฐ Other Tools

๐Ÿง  AI Components

๐Ÿ”ฎ Large Language Models

For Pre-production: OLMO2:13B

A state-of-the-art, fully open-source language model by the Allen Institute for AI, designed for transparency and performance.

  • Fully Open Ecosystem
  • High Performance: Trained on 5T tokens, outperforming Llama-3.1 8B and Qwen 2.5 7B in academic benchmarks.
  • Flexible Use: Supports text generation, reasoning, and fine-tuning; optimized for single-GPU inference.

For Production: LLaMA3.2:latest

A cutting-edge, multimodal model by Meta, optimized for agentic workflows and code-driven applications. Designed for developers building autonomous AI systems and coding tools.

  • Agentic Task Execution
  • Built-in Tool Calling: Interacts dynamically with external APIs/tools (e.g., Brave Search, Wolfram Alpha) for real-time data, code execution, and query solving.
  • Workflow Automation: Enables multi-step reasoning, parallel tool execution, and iterative problem-solving (e.g., analyzing weather data, synthesizing answers).

DeepSeek-R1:8B

A high-efficiency, open-source model by DeepSeek AI, specialized in reasoning and code generation.

  • โœ… Advanced Reasoning & Code Generation
    Excels at step-by-step problem-solving in coding and math. Outperforms GPT-4o and Claude-3.5 on LiveCodeBench (65.9% pass@1) and MATH-500 (97.3% pass@1).
  • โœ… Cost-Efficient Architecture
    Trained under $6M using FP8 precision and multi-token prediction, cutting memory by 75% while preserving accuracy. Distilled from a 671B MoE model with strong performance (1691 CodeForces rating for 32B distilled version).
  • โœ… Open & Customizable
    Apache 2.0 license for commercial use. Weights and tools available on Hugging Face. Supports fine-tuning via PyTorch/HuggingFace for specialized domains (e.g., code optimization, technical docs).
  • โœ… Multilingual & Scalable
    Maintains reasoning in French. Quantized versions support edge deployment on consumer GPUs.
  • โœ… Transparent Workflow
    Reveals logical steps and thought processes, supporting error tracking and logic auditing.

๐Ÿ—„๏ธ Vector Databases

Qdrant

๐Ÿ”ค Embedding Models

dangvantuan/sentence-camembert-large

sentence-transformers/all-MiniLM-L6-v2

๐ŸŽญ Orchestration Framework

Langchain

๐Ÿค– Agent Framework

LangGraph

๐Ÿ“š RAG Implementation

๐Ÿ” Retrieval Pipeline

Retrieval Pipeline Diagram

๐Ÿ“„ Document Processing

๐Ÿ’พ Knowledge Base Management

๐Ÿ’ฐ Financial Data Integration

๐Ÿ“Š Data Sources

Jibaya.tn : The official online portal of the Tunisian tax administration, managed by the Direction Gรฉnรฉrale des Impรดts (DGI). It offers a range of digital services for taxpayers, including online tax declarations, payment of taxes, and access to comprehensive fiscal documentation.

๐Ÿ“ˆ Market Data APIs

๐Ÿ“‰ Financial Analytics Tools

๐Ÿ”’ Security and Compliance

๐Ÿ›ก๏ธ LLM Output Safety

๐Ÿ” Financial Data Protection

โš–๏ธ Regulatory Compliance

๐Ÿ“ Evaluation Framework

๐ŸŽฏ RAG Quality Metrics

๐Ÿ“Š Agent Performance Metrics

โœ… Financial Advice Accuracy

๐Ÿš€ Deployment

โ˜๏ธ Infrastructure

The project runs on an NVIDIA DGX A100 server equipped with a single 80GB GPU, providing high memory bandwidth and compute power optimized for AI workloads. This infrastructure enables efficient training and inference for large-scale deep learning models.

๐Ÿ“ก Monitoring

๐Ÿ“ˆ Scaling Strategy

๐Ÿ‘ฎ Model Governance

๐Ÿท๏ธ Versioning

๐Ÿงช Training Data

โš–๏ธ Bias Mitigation

๐Ÿ“ Directory Structure

๐Ÿš€ Getting Started

  1. Clone the repository
    git clone https://github.com/fouratmansouri/Auditron.git && Auditron
  2. Create and activate a virtual environment
    On macOS/Linux:
    python3 -m venv venv && source venv/bin/activate
    On Windows:
    python -m venv venv && .\venv\Scripts\activate

๐Ÿ™ Acknowledgments

This project was developed in partnership with Esprit School of Engineering and Esprit's financial department.

Special thanks to:

About

This project, Auditron, was created as part of the Challenge-Based Learning (CBL) AI project at Esprit School of Engineering - Tunisia.

Topics

Resources

Stars

Watchers

Forks

Contributors 4

  •  
  •  
  •  
  •