ABOUT THE CLIENT
Finotor is a fintech startup founded in 2008. The company offers a financial management platform that integrates artificial intelligence to automate and optimize business financial processes.
CHALLENGE
The project presented two main challenges: data preparation and AI model optimization. Financial data came from multiple sources with non-standardized formats, inconsistent amounts and dates, and incomplete histories. The system had to ensure accurate real-time processing while complying with financial industry standards.
DATA TYPE
Our system processes raw financial transactions, composed of different elements that need to be analyzed and classified.
Input Data Transactions appear as bank lines containing:
A non-standardized bank description
An amount
A transaction date
Banking information (BIC, IBAN, etc.)
Elements to predict For each transaction, we need to identify four key pieces of information:
Transaction Type This involves classifying the nature of the operation: transfer, direct debit, card payment. This information is usually present in the description but in different forms depending on the banks.
Category We need to determine the purpose of the transaction: salary, rent, supplies, etc. This categorization must be done by analyzing the complete context of the transaction, as the description alone is not always sufficient.
Client and Supplier ID For each transaction, we need to identify:
The client entity making or receiving the payment
The relevant supplier entity These identifications require matching bank descriptions, often poorly standardized, with our entity database.
Our main challenge lies in the non-standardization of banking data and the need to understand the context of each transaction for accurate classification.
SOLUTION
Our solution relies on three main features:
Automatic bank reconciliation intelligently matches transactions and detects anomalies in real-time, significantly reducing processing errors.
Predictive analysis generates recommendations based on historical data and enables KPI tracking through dashboards.
Cash management optimizes financial flows with real-time alerts and automated suggestions.
PROCESS
Our development was structured in three stages:
Data Preparation
Development of a cleaning pipeline to standardize formats and data
Management of missing values and unification of categorizations
Rigorous validation to ensure financial data integrity
Model Training
Comparative tests between LLM, XGBoost and neural network
Final selection of neural networks for their better performance
Continuous hyperparameter optimization
Testing Phase
Cross-validation on representative datasets
Real-condition testing with specific financial metrics
Iterative adjustments based on experience feedback
METHODOLOGY
Our team combines various essential expertise:
AI Engineers specialized in Machine Learning
Product Owner
Lead developer
Development took place in three phases:
An exploration phase allowed comparison of different AI approaches and validated the superiority of neural networks for our use case.
Iterative development allowed progressive refinement of system performance in real conditions.
Deployment was done gradually to ensure good user adoption.
MODEL COMPARISON
Our comparative analysis evaluated three approaches: neural networks, XGBoost, and LLM. Neural networks clearly stood out with superior performance on all key aspects:
For transaction type classification, neural networks achieve 98.3% accuracy with only 0.22% uncertain cases, outperforming LLM (92%) and XGBoost (96%).
On category identification, which is more complex, our neural networks maintain 84.44% accuracy, compared to 80% for XGBoost and only 59% for LLM. Excluding uncertain cases (confidence threshold >65%), accuracy rises to 92.9%.
Client and supplier ID identification represented the biggest challenge. Our neural networks achieve 76.51% and 77.30% accuracy respectively, far ahead of XGBoost (65% and 48%) and LLM (6.67% and 31.65%).
A crucial point: our neural networks generate significantly fewer uncertain cases than XGBoost while maintaining superior accuracy. This combination of high accuracy and low uncertainty is ideal for a financial application requiring reliability and comprehensiveness.
This thorough analysis confirms the superiority of neural networks for our use case, offering the best compromise between accuracy, certainty, and coverage of different classification aspects.
RESULTS
The system achieves 98% accuracy in its predictions and processes over 90,000 transactions. Automation covers 85% of financial processes with a 75% reduction in errors.
PERSPECTIVES
We are currently working on integrating a chatbot for customer support and developing new automation features to enhance the platform.