Below are some projects where I collaborate
- AI in supply chain (TESCO)
- Insurance risk models – DL
- Cybersecurity (Santander)
- Fraud Detection
- Credit Risk
- Bayesian Optimisation
- Retargeting advertising advertising optimisation
- Digital Marketing optimisation.
Retargeting advertising optimization
Mainad works on retargeting advertising. They use Google Cloud to serve thousands of ads per second and up to 100 million ads every day worldwide with a revenue of tens of millions USD.
They want to optimize the efficiency of the campaigns by targeting and pricing properly the right users in order to have maximum traffic in the most cost effective way.
We develop a solution based on LSTM neural networks to predict clicks and sales based on past behaviour of million of users. The solution was tested and deployed in the Google Cloud pipeline and is currently in production.
We were able to increase CTR by more than 20% and sales by 18% under the same budget.
Medical image processing
Contextvision is a major provider of image processing software throughout the world. They pioneered solutions for X-ray and Ultrasound image filtering and enhancement.
Contextvision wanted to develop a solution for an ultrasound probe to automatically segment, in real-time, arteries and veins in arms and neck captured in ultrasound video.
We develop a solution based on Deep Neural Networks using a convolutional LSTM (CNN+LSTM) to automatically identify arteries and veins in a very noisy video with a speed of up to 10 frames per second (without reliance on GPU). The code is available on http://github.com/farquasar1/CONVLSTM
We were able to detect arteries and veins with a consistent dice score of up to 80%. The algorithm was tested and deployed in a US client (Bard).
DirectLine group is a major insurance group that handles more than 3 million policies with an annual revenue of up to £2 billion.
DirectLine required to develop a new pricing algorithm based on the latest machine learning algorithms to better price and underwrite auto policies.
We develop a solution based on Lightgbm and Xgboost machines, including several techniques to handle high cardinality categorical data and a new method that does not rely on modeling risk as a Poisson process.
We delivered a set of algorithms that, for extreme losses, like body injury excess, increased the Gini index by more than 20% while keeping the same binarized Mean Average Error on equal exposure bins.
Tesco stock management
Tesco is the major distribution chain in UK with a turnover well above £32 billion with a complex supply chain and stock management system.
The client challenged us to develop an algorithm to estimate the products on shelves and respective uncertainty – due to losses, waste and errors in the supply chain.
We develop an algorithm based on Dynamical Bayesian Networks to estimate the stock levels at each point in time until the next MST (Modern Stock Take) takes place – every 6 months.
The model was able to detect if lack of sales was originated in no demand or no products on the shelf. An estimate of saving around £80 million was achieved through reduction of stock routines and sales availability improvement.
See the full publication here