Jean-Michel Nairac

OVERVIEW
A bank client engaged Ciel Finance Data Services Ltd to automate its loan application process, which was manual and tedious. Loans were pre-approved based on a simple checklist, and final approval was a qualitative exercise reviewed by multiple committees.

SOLUTION
1. Design a new loan application process.
2. Build ML models to produce a credit score for loan approval.

I designed a new data pipeline that involved:
• An ETL process of the data into Microsoft Azure.
• A Python script to execute the ML models and predict the credit scores.
• A SQL database to store the results and push them to the client via email.
• Azure Data Factory to orchestrate and automate the process.

I built two ML models and combined them to produce a credit score. One model predicts the probability that the committees will reject the application, and the other predicts the probability of default given that the committees approved the loans. Each model uses deep learning (TensorFlow and Keras), and gradient boosted trees (LightGBM and XGBoost). Features were engineered with statistical methods using Pandas, SciPy, and NumPy.

OUTCOME
We saved the bank a lot of time by discontinuing several committees and improving its loan default rate.

OVERVIEW
An anomaly detection tool I built to detect anomalies in bank transaction data. The tool is an ensemble machine learning model of many models with various hyperparameters of KNN, HBOS, Isolation Forest, PCA, and COPOD models.

CHALLENGE
The client did not have a system to detect anomalies, so they used only ad hoc samples.

SOLUTION
I engineered research-backed features for anomaly detection in bank transaction data. The data pipeline involves an ETL process of incrementally loading data into Microsoft Azure SQL Database using Azure Data Factory. A Python script containing the anomaly detection tool is automatically executed on the incrementally loaded data, and a CSV file containing anomalies is sent directly to the client by email.

OUTCOME
The client is using my anomaly detection tool because:
• It can run all transaction data as opposed to a sample.
• It automatically filters outliers, which is a vast improvement over the manual process.
• It provides an outlier score so that the client can prioritize the investigation of potential anomalies as they see fit.

The client confirmed that the tool is effective in detecting more data quality issues and fraudulent activities than the ad hoc sampling process.

OVERVIEW
This was a Kaggle competition to predict volatility for options trading.

CHALLENGE
We received trade data and order book data in 10-minute intervals per second, totaling over one billion data points. Our task was to predict 10-minute volatility with performance assessed using root mean squared percentage error (RMSPE) between true and predicted volatility.

SOLUTION
I performed well in this competition because I:
• Applied data analytics to correctly handle errors, such as trading halts, stock splits, mergers, and low trading volumes.
• Used the correct cross-validation technique (StratifiedGroupKFold).
• Engineered research-backed features (papers on algorithmic trading are freely available on Google Scholar and arXiv).
• Re-engineered timestamps and captured correlation between timestamps.
• Engineered features that correctly captured the correlation between stocks (Pandas and NumPy).
• Used ensemble methods of linear models (scikit-learn), gradient boosted trees (LightGBM), and deep learning (TensorFlow and Keras).

OVERVIEW
The client manually recorded electricity meter readings. This tedious process consumed a lot of the client's resources because many meter readings had to be recorded on a regular basis.

PROPOSAL
We proposed using security cameras to capture images and automatically extract the meter readings using artificial intelligence and machine learning.

SOLUTION
We used Microsoft Azure tools to build the solution and the custom vision software to detect objects—meter screens from security image captures. Microsoft also provides Computer Vision OCR for text extraction from images. We combined these into a SaaS solution to detect meter readings from security images.

The pipeline is engineered as follows:
1. The security image captures are uploaded to Azure Blob storage.
2. This triggers a Python script that makes an API call to the custom vision predictions and another API call to the Computer Vision OCR.
3. Meter readings are extracted and pushed to an Azure SQL database.

OUTCOME
The client saves a lot of time, and resources are freed up to focus on other tasks.