About

Brief Summary:

I am currently a Computer Vision Research Engineer at Zebra Technologies in London, UK. My main achievements are - improving model predictive performance, actively contributing to and maintaining our internal core repository, and greatly reducing the latency across the company’s models when deployed on edge devices with NPUs. So far, my contributions on improving predictive performance and latency of vision models have been algorithmic and have affected all models in our business unit, thus making them high impact contributions.

I graduated from MSc Machine Learning at University College London. With experience and interests in Reinforcement Learning and Generative Models. I have also completed courseworks related to NLP (GAN-BERT for semi-supervised learning) and Computer Vision (VAE, Neural Style Transfer).

Before this I was working as an Applied Scientist at NewDay, where I applied Deep Learning (Natural Language Processing) for text augmentation and sentiment analysis. I also worked on improving the performance of an incumbent chatbot model. Other projects I worked on include building a churn prediction model and marketing channel analysis using Bayesian models.

I did my bachelor’s degree - BSc Statistics, Economics and Finance (accredited by the Royal Statistical Society) - at University College London.

Career interests:

So far my contributions have been general and high impact, I am flexible when it comes to optimising predictive and speed performance of models of different modalities. My interests lie in Research Engineering, therefore. See my CV for more information.

Open source contributions:

• PyTorch contribution - Add a strict kwarg and fix a bug in nn.Module.set_submodule.
• PyTorch Lightning contribution - fix a bug in callbacks.ModelCheckpoint.

Featured projects:

• Reinforcement Learning algorithms - such as PPO and DQN.
• Instruction fine-tuned GPT-2 using my own implementation of LoRA with PyTorch and Lightning. Trained in bfloat16 and deployed to a Gradio app with Docker.
• Fine-tuned BERT with LoRA in bfloat16 for text classification - Using PyTorch, Lightning and Hugging Face. Local testing with Flask and Docker and deployed to a Gradio app with Docker.