# Luke Byrne 📍 Glasgow, Scotland | ✉️ [[email protected]](mailto:[email protected]) | 🐙 [github.com/luke-byrne-eng](https://github.com/luke-byrne-eng) | 💼 [linkedin.com/in/luke-byrne-engineer](https://www.linkedin.com/in/luke-byrne-engineer) ## About AI / Machine Learning Engineer specialising in safety-critical AI for data-scarce domains. Delivered applied AI systems for the European and UK Space Agencies, including a family of multi-modal transformers integrating several sensor types (vision, time series, depth, radar etc.). Industry background in robotics integration and embedded electronics design & manufacture. Research outputs on transformer design, dataset creation, data augmentation, and probabilistic inference. Expert in overcoming data scarcity through self-supervised pretraining of foundation models. ## Technical Skills * **Machine Learning & Computer Vision:** Deep learning, vision transformers, CNNs, self-supervised learning (MAE), foundation models, physics-informed neural networks (PINNs), RAG, vector databases (ChromaDB), OpenCV, scikit-learn, LoRA/QLoRA * **Uncertainty Estimation:** Hamiltonian Monte Carlo, Bayesian neural networks, Monte Carlo dropout, uncertainty estimation, variational inference * **ML Engineering & Deployment:** PyTorch, HuggingFace (transformers, datasets, PEFT), Docker, REST APIs, distributed training, mixed precision (AMP), model quantization (INT8, bitsandbytes), ONNX, local LLM serving (vLLM), Weights & Biases, Tensorboard, Optuna, SQL (SQLite) * **Robotics & Signal Processing:** ROS, SLAM (monocular depth, Time of Flight, LiDAR), Kalman filters, Fourier analysis, PLC & industrial robot programming (Fanuc, ABB) * **Programming & Tools:** Python, C++, MATLAB, Linux, Git ## Commercial Experience **University of Strathclyde**, Research Assistant (Client: European Space Agency) | *2025 – 2026* *(Part-time commercial contract undertaken concurrently with PhD)* * Delivered a production-ready physics-informed multi-modal transformer to the European Space Agency for integration into the IOTA space simulator, with comprehensive unit tests and code-reviewed delivery * Designed an architecture integrating Light Curves, Hyperspectral, Radar, ISAR, and Laser Ranging measurements, capable of processing sequences from minutes to hours in duration * Quantified rotation rate and attitude estimation uncertainty via Monte Carlo inference, providing calibrated confidence intervals for mission-critical decision-making * Authored the ML methodology and results sections of the ESA technical deliverable **University of Strathclyde**, Research Assistant (Client: UK Space Agency) | *2024 – 2025* *(Full-time commercial contract; PhD paused for sabbatical)* * Delivered a Fourier-domain model family to UK Space Agency stakeholders estimating space-debris rotation rates from hyperspectral light curves, achieving <0.5 degrees/second RMS error * Implemented Hamiltonian Monte Carlo sampling for calibrated uncertainty estimates, supporting mission-critical space situational awareness applications * Contributed the ML methodology and results sections of the UKSA technical deliverable **PS Autogrinding**, Robotics Engineer | *2022* * Integrated computer vision for real-time part identification and tracking on conveyor systems * Programmed Fanuc and ABB industrial robots for pick-and-place operations * Assembled robotic loading cells and electrical control cabinets **GM Instruments**, Electronics Engineer | *2021 – 2022* * Designed, manufactured, and calibrated medical signal processing devices, ensuring compliance with strict medical testing standards * Performed component-level failure analysis and liaised directly with customers ## PhD Research **University of Strathclyde**, PhD Researcher | *2022 – 2026* * Pretrained a vision transformer (ViT) foundation model using Masked AutoEncoding (MAE) on 3 months of unlabelled dairy-barn video, establishing a scalable, annotation-free pipeline for commercial Precision Livestock Farming * Finetuned the foundation model to >90% accuracy classifying complex behaviours from only 130 labelled 10-second clips, demonstrating exceptional data efficiency; exported to ONNX for prototype on-farm deployment with the industrial sponsor * Engineered and open-sourced [AVA-Bovine](https://github.com/Luke-Byrne-Eng/ava-bovine), a unified benchmark for domain generalisation between dairy barns, accelerating development of robust models capable of zero-shot deployment across diverse environments * Designed novel activation functions (SinGLU; TMLR submission) and a frequency-aware data augmentation method (SFC-Conv) improving vision transformer efficiency for resource-constrained deployments and out-of-distribution generalisation * Implemented Monte Carlo dropout for out-of-distribution detection, enabling safer human-in-the-loop AI through estimates of epistemic uncertainty * Managed end-to-end MLOps and large-scale training pipelines on the ARCHIE-WeSt HPC cluster (NVIDIA A100), using Automatic Mixed Precision and Bayesian hyperparameter optimisation ## Engineering Projects * **Self-hosted Research LLM Platform** — Deployed a locally hosted, secure AI research assistant (Open WebUI, vLLM, Qwen) with a custom RAG pipeline (ChromaDB, reranking, Docker-deployed REST API). Configured OpenAI- and Anthropic-compatible API endpoints for integration with existing tooling (Claude Code), keeping sensitive research data off third-party APIs. * **Voice-Cloned Audiobook Pipeline** — Fine-tuned open-source TTS models with QLoRA on text-audio pairs extracted from video game dialogue, building an end-to-end audiobook generation pipeline from arbitrary text inputs. ## Education **PhD, Artificial Intelligence**, University of Strathclyde | *2022 – 2026 (Expected)* *Funded by **EPSRC** and **Peacock Technology Ltd*** **MEng, Electronic & Electrical Engineering**, Glasgow Caledonian University | *2020 – 2021* *(Mechatronic & Robotic Engineering stream)* **BEng (Hons), Electronic & Electrical Engineering**, Glasgow Caledonian University | *2017 – 2020* *First-Class Honours* ## Publications and Presentations * **L. Byrne**, P. Murray, "SinGLU: A Systematic Study of Gated Linear Unit Design Space Reveals Viability of Sinusoidal Activations in Small Vision Transformers", **Transactions on Machine Learning Research (TMLR)**, *under review* * **L. Byrne**, P. Murray, "AVA-Bovine: A Benchmark of Barn-to-Barn Domain Generalisation in Cattle Behaviour Recognition", **Computers and Electronics in Agriculture**, *under review* * **L. Byrne**, B. Lee, A. Ulrichsen, P. Murray, M. Marshall, "Binary classification of dairy cow behaviour using vision transformers", **International Society for Applied Ethology (ISAE)**, Tallinn, Estonia, 2023 — *oral presentation* * **L. Byrne**, P. Murray, M. Marshall, "SFC-Conv: Spatial, Frequency, and Colour Selective Convolutions for Neural Network Data Augmentation", *in preparation*