About Me
Hi, my name is Thomas (Tom) Bale, and I'm a student, developer, business owner, and Ironman AG Athlete. I am pursuing a Bachelor's degree in Computer Science at the University of Bristol, currently achieving 78.33% (First Class) average. I serve as Treasurer and Planning & Control Team Lead in Formula Student AI, and I am the Founder, President, and Competitions Lead of UoB Quantum Computing Society. Alongside my university work, I explore my strong super-curricular passions for quantum computing, AI, and HPC. I also participate in triathlons/Ironman training and co-own a performance cycling products small business.
My academic, super-curricular, and professional experience have given me a strong foundation in full-stack development, computer vision, ML, quantum computing, and algorithms. Find my projects below or on my GitHub relating to quantum and AI.
Please contact me below if you have any queries, or find out more about me on my LinkedIn profile!
My projects
Quantum cross-chain arbitrage
Winners - ETH Oxford DeFi hackathon prize ($5000)Vyperlang bounty - 2nd placeQuantum-Enhanced Cross-Chain Arbitrage Bot (QXAB) uses Quantum Approximate Optimisation Algorithm (QAOA) and Flare's blockchain protocols to execute fast, secure, and profitable cross-chain arbitrage trades via flash loans.
- Blockchain
- Quantum
- Vyper
- QAOA
Llama 8B @ ISC 2025
ISC25 Student Cluster Competition: LLaMA Fine-Tuning Task Performance Optimisation Techniques: To maximise training throughput on 8× H100 GPUs, we employed: FlashAttention 3 for H100-specific kernel optimisations, significantly reducing time per step while maintaining numerical stability. Transformer Engine with FP8 support for reduced memory usage and improved compute efficiency via autocasting and fused operations. FP8 precision for training to reduce memory footprint and accelerate computation, chosen over BF16 due to superior performance on H100s. Effective batch size of 256 (per_device_batch_size=16, num_gpus=8, gradient_accumulation_steps=2) for stable convergence within memory constraints. Accuracy Optimisation Techniques: DoRA (Delta-Orthogonal Rank Adaptation) - a fine-tuning technique improving upon LoRA and qLoRA by maintaining full-rank parameter contributions with minimal training overhead, achieving higher validation accuracy. Trained for 5 epochs with max_steps=69 per epoch, covering the dataset within time budget while ensuring convergence without overfitting. Key Decisions: Chose DoRA over qLoRA prioritising model quality over minimal resource use given 8× H100 availability. Integrated FlashAttention 3 and Transformer Engine for significant throughput gains despite complexity. Selected FP8 over BF16 for larger batch sizes and faster execution, with Transformer Engine ensuring reliable training convergence.
- Python
- PyTorch
- HPC
- CUDA
- FP8
- FlashAttention
- Transformer Engine
- DoRA
- LLM
- Fine-tuning
YukiGPT
A decoder-only GPT implementation based on the transformer architecture proposed in 'Attention is All You Need', without encoder or cross-attention components. Built from scratch in Python following Andrej Karpathy's educational series. Trained to generate F1 radio messages. Includes implementations of: bigram language model, self-attention mechanism, and dataset utilities. Currently working towards expanding the model and implementing a custom tokeniser.
- GPT
- Transformers
- NLP
- Python
- PyTorch
- Deep Learning
- Attention Mechanisms
F1 Ghost Car
An overhead reconstruction visualisation tool for Formula 1 qualifying sessions. Creates animated comparisons of two drivers' fastest qualifying laps for a given Grand Prix. Features include: animated car tracking with realistic track layouts, follow camera mode that tracks the cars, optional realistic track surface rendering, and support for comparing any two drivers from 2018-2024. Built in Python using matplotlib for visualisation and F1 position data for accurate lap reconstruction.
- F1
- Data Visualisation
- Matplotlib
- Animation
- Sports Analytics
- Python
Parallel vs Distributed Implementations for Conway's Game of Life
86% scoring coursework. Implemented, optimised (including halo exchange, communication overhead, architecture considerations) and compared parallel (concurrent go programming) and distributed (AWS EC2 instances using RPC calls) versions for Conway's Game of Life. Report includes benchmarking algorithmic performance, Matlab for graphs and identifying bottlenecks using CPU profiling. Optimised network communication via publish/subscribe model ensuring fault tolerance and scalability.
- Go
- AWS
- Academic writing
- LaTeX
- Concurrent programming
- Distributed
- Algorithmic optimisations
AI for chess in 3 Dimensions
The first ever game and engine for Chess in 3 dimensions. Makes use of an AI with a NN and Mini-Max. Includes a customisable UI. Includes and extensive report, with: Code, UML diagram, Flow charts for complicated algos and AI, Unity environment, configuration and assets, AI justifications, Client feedback and interviews, Research. Maximum mark scoring project.
- C#
- Game development
- Unity
- OOP
- AI
- ML
- MNs
Advent of Code
My advent of code daily solutions. All 2024 days solved in python. 2025 solved in a different language each day.
- Python
- algorithms
- graph theory
- Number Theory
- Dynamic Programming
georgia-AI
An AI designed for inferring the NYT Wordle without playing the game, only using the results of how up to 20 other people had guessed. Implements NN and beam search with sampling.
- Python
- NumPy
- Pandas
- Torch
- NNs
Other Projects - GitHub
These include: Quantum hackathon for UoB; ML challenge for sixth form society; Algorithms II minigame; .PGM to .SK converter; mandelbrot set visualisation; and decimal to binary converter.
- C
- Haskell
- Python
- Game development
- Hackathons
My Experience
Demonstrator and Graduate Teacher
University of Bristol
Aug 2025 –
Delivering Lectures and Workshop Exercises. Mentoring a small team of students through their year-long project (Software Engineering Project). Helping students in workshops (Computer Systems A, Programming Languages and Computation).
UK HPC Student Team
UKSCC
May 2025 – Jun 2025
Represented the UK at ISC SCC. Optimised OpenMX and LLMs (llama) on a 208 core, 8xH100 cluster. Experience with load balancing, networking, OpenMP, MPI, and CUDA programming. Worked with SLURM, LAPACK, BLAS, ScaLAPACK, FFTW, ELPA, OpenBLAS, Intel MKL, and custom profiling. Explored LoRA, QORA, DORA, transformer engine, and FP8, flash attention 3 for llama 8B.
Machine Learning Research Assistant
University of Bristol
Feb 2025 – Aug 2025
Developed a scalable ML workflow for generating photorealistic, emotional faces for psychological research.
Co-Founder & Operator
Veloworks Components
Sep 2024 –
Co-founded and operate a business producing 3D-printed performance cycling components. Manage operations encompassing sponsorship, marketing, technical stages, and financial activities.
Machine Learning Software Engineer Intern
DigitalU3
Sep 2024 – Mar 2025
Engineered a machine learning-based system with a strong focus on efficiency and scalability. Collaborated in an AGILE team, contributing in sprints, code reviews, and iterative delivery. Developed a scalable web application with integrated backend and structured database.
Demonstrator and Graduate Teacher
University of Bristol
Aug 2025 –
Delivering Lectures and Workshop Exercises. Mentoring a small team of students through their year-long project (Software Engineering Project). Helping students in workshops (Computer Systems A, Programming Languages and Computation).
UK HPC Student Team
UKSCC
May 2025 – Jun 2025
Represented the UK at ISC SCC. Optimised OpenMX and LLMs (llama) on a 208 core, 8xH100 cluster. Experience with load balancing, networking, OpenMP, MPI, and CUDA programming. Worked with SLURM, LAPACK, BLAS, ScaLAPACK, FFTW, ELPA, OpenBLAS, Intel MKL, and custom profiling. Explored LoRA, QORA, DORA, transformer engine, and FP8, flash attention 3 for llama 8B.
Machine Learning Research Assistant
University of Bristol
Feb 2025 – Aug 2025
Developed a scalable ML workflow for generating photorealistic, emotional faces for psychological research.
Co-Founder & Operator
Veloworks Components
Sep 2024 –
Co-founded and operate a business producing 3D-printed performance cycling components. Manage operations encompassing sponsorship, marketing, technical stages, and financial activities.
Machine Learning Software Engineer Intern
DigitalU3
Sep 2024 – Mar 2025
Engineered a machine learning-based system with a strong focus on efficiency and scalability. Collaborated in an AGILE team, contributing in sprints, code reviews, and iterative delivery. Developed a scalable web application with integrated backend and structured database.
My Education
BSc in Computer Science
University of Bristol
Sep 2023 – Current
78.33% avg; First Class Treasurer and Planning & Control Team Lead in Formula Student AI. Founder, President, Competitions Lead UoB Quantum Computing Society.
A-Levels
Colchester Royal Grammar School
Sep 2021 – Jul 2023
A*A*AA - Computer Science (ranked 1st in cohort; 100% NEA), Maths, Further maths, Physics.
BSc in Computer Science
University of Bristol
Sep 2023 – Current
78.33% avg; First Class Treasurer and Planning & Control Team Lead in Formula Student AI. Founder, President, Competitions Lead UoB Quantum Computing Society.
A-Levels
Colchester Royal Grammar School
Sep 2021 – Jul 2023
A*A*AA - Computer Science (ranked 1st in cohort; 100% NEA), Maths, Further maths, Physics.
My skills
- Python
- Go
- Java
- C
- C#
- Haskell
- HTML/CSS
- JavaScript
- SQL
- API/REST
- Git/Github
- React Native
- PyTorch
- TensorFlow
- Qiskit
- AWS
- Cursor
- AGILE & Test-Driven Development
- Quantum Programming
- Machine Learning (ML)
- Artificial Intelligence (AI)
- High performance computing (HPC)
- CAD
- Generative AI
- Computer Vision
- LLM
- A/B Testing
- NLP
My achievements and interests
Competitive Swimmer
Competing at BUCS and swimming for Bristol University performance squad.
Ironman Switzerland - 3.8km Swim, 180km Bike, Marathon Run
Completed in 10h30m after 9 months of training with peak weeks averaging 20-25h/week with 10,000m swimming, 500km cycling and 70km of running.
Climbing
Enjoy indoor bouldering, climbing v6-7. If I'm not on my laptop or in the pool, you'll probably find me in a climbing gym!
Formula Student AI
Treasurer and Planning & Control Team Lead in Formula Student AI at University of Bristol.
Formula 1 Projects
Developed F1-related projects including F1 Ghost Car visualisation tool for comparing qualifying laps, and various other projects working with F1 data.
High Performance Computing
I have represented the UK at ISC25 Student Cluster Competition and represented the University of Bristol at CIUK high performance computing cluster challenge twice.
Quantum Computing
I have casually studied quantum computing for 3 years. Founded University of Bristol Quantum Computing after having a large uptake in applications for a UoB hackathon team for MIT iQuHack. This first involved leading a team to participate in a quantum computing hackathon hosted by MIT iQuse. Now, I run weekly sessions and plan to provide training and mentorship for new students to the area as well as designing and hosting accessible intra uni competitions.
Blind Solve Rubik's Cube
I all enjoy all types of puzzles, from Rubik's cubes to jigsaw puzzles to escape rooms. At the moment, I want to get sub 4m for a blind 3x3 Rubik's cube solve.
Contact me
Please contact me directly at tokbale@outlook.com