Machine Learning Developer in United Kingdom

I developed a system to find the optimal distillation sequence and parameters for a specified non-ideal mixture. This dramatically reduces the time it takes for a process engineer to design a system that meets the requirements.

I developed a deep learning pipeline for an application that required pixel-perfect segmentation masks for retail. It involved carefully selecting and combining state-of-the-art models to achieve the required performance.

I contributed to an industry disruptive startup that aims to solve drug discovery with AI. I worked on software engineering and DevOps for large-scale cluster systems to support the research team and designed and implemented a system architecture that made experimentation more cost-efficient and faster. Also, I proposed architectural changes to the research team regarding generative models.

I worked on a proof of concept for a financial forecasting model that involved large amounts of unevenly spaced time series data. I developed a PyTorch data loader interface to an SQL data warehouse to efficiently train models at scale and tailored deep learning models.

I worked on a machine learning technology demonstration that turned sketches into pieces of art. I built on top of cutting-edge deep learning architectures using generative adversarial networks to create a system that can hallucinate artistic paintings based on the input edges. I used Python with Tensorflow to build the architecture, implement the data pipeline and facilitate distributed training. Since the work that we've done for clients was confidential, this demo could showcase perfectly the capability of the team, and this demo, in particular, had a significant impact on our presence in the AI consulting domain.

This technology demonstration aimed to reconstruct distorted images, so the model not only did pixelwise mapping, but it also had to fill in missing information based on the context. Generative adversarial networks are excellent in doing this, and it was shown how the model can change the output depending on the environment, I used Python with Pytorch for the training and a Flask back-end with a Typescript front-end for the deployed system.

I worked on a non-linear controller that is able to learn the system dynamics that it controls. An artificial neural network is trained online to predict the system states to provide better performance than a simple PID algorithm. It was applied to a simple 2 degrees-of-freedom robotic arm to prove the concept. I developed the mechanics and the electronics for the prototype, simulated the dynamics in C++, and deployed the system to a low-cost Atmel microcontroller using Embedded C.

I developed system design for a humanoid walking robot, simulated and prototyped the control algorithm using Simulink. I implemented the Linux middleware for the actuators and sensors on the onboard controller.

I worked on the simulation environment for the IMAV drone competition. I used Gazebo, ROS, and Simulink to provide a high-fidelity simulation that can be used to evaluate autonomous flight controllers.

I developed an RL agent for cryptocurrency trading that uses OHLCV data and technical indicators. The model was trained and backtested with data from the last few years, and it showed promising results, outperforming baseline strategies.

I evaluated different RL algorithms (PPO, IMPALA) and model architectures (fully connected, convolutional, attention-based) to find the best performing agent.

I developed the model using PyTorch and Ray, which enables massive scaling to reduce the training time significantly. For backtesting and live trading, I used Freqtrade.