Michael Gallimore

Experienced at training a wide variety of machine learning models on large datasets including:

• audio
• video segmentation and classification
• LLMs
• tabular data for insurance premium prediction and collaborative filtering (recommender systems) 
• and more

In my last two roles I delivered significant improvements to previous approaches used for receipt scan classification (NLP) and birdsong detection (audio) in two very different industies. 

I've spent 5 years training machine learning models, and I'm passionate about finding ways to transform datasets into predictions by applying the skills I've gained. 

I'm flexible, a great communicator and easy to get along with. 

Some of my skills are listed below:

• Audio, image and Video classification models
• Segmentation models for video or images
• LLM training, finetuning and RAG systems
• Delivering bespoke systems tailored to the needs of the project
• Deployment using Docker, Dagster.
• Reccomender systems
• Industry experience in Acoustics, Electronics, Geology, Conservation / birdsong detection.

I have completed two work contracts in machine learning - in one I was tasked with building a birdsong detector. I used my knowledge of ML principles to pick training audio from within the correct domain, used hard example mining and assisted learning, developed methods to show the precision, recall, F1 score and accuracy on a series of validation and test sets, and to compare against other SOTA modles. I delivered an increase in accuracy from 86 to 95% compared with the best SOTA model for the species I was tasked with detecting. 

In my first work contract, I was tasked with making a classification system for text from receipts, which would auto-populate fields in an expense tracking app. I delivered an increase in classification accuracy from 40% to 80% compared with the original system implemented at the company. The text database I had access to was noisy, since the OCR which had been used wasn't perfect. I developed a system to clean out any noise from the signal, then built a naive bayes based method for classifying the receipt contents. The model was intended to work on edge devices with no internet connection, so I delivered the model's output in a dictionary format for quick lookup. 

Both of these roles were remote, contract based work where I communicated about the needs of the project with people from different industries to deliver a successful result. 

I completed a BSc Hons degree in Acoustics in 2011 where I learned computer science, physics, maths, electronics and other disciplines. This is where I first encountered coding in C++ and C#. We implemented genetic algorithms for searching a problem space, and made digital signal processing systems and physics simulations in MATLAB. 

Since graduating University I have taught myself python, full stack web development, and learned to train Deep Learning models from textbooks including the FastAI course Deep learning for Coders.