A. K. Mackay
Contents

Andrew K. Mackay

Machine Learning & AI Engineer · Zaragoza, Spain

StatusWorking as an AI Engineer at MediaRadar.

Abstract. I am a machine learning engineer with a research background in reinforcement learning and robotics. I hold an MSc in Computer Science from the University of Innsbruck, have published on reinforcement learning for robot navigation, and have put models in front of real users, from eye-tracking diagnostics validated with over 4,000 patients to LLM solutions for industry clients.

I now work as an AI Engineer at MediaRadar, building machine learning pipelines that classify media content. Before that, my work was on LLMs, retrieval and agents, where I care as much about measuring a system as about building it.

Selected work

2026 – present

Ultralytics YOLO11, YOLO-World, PyTorch, OpenCV, Google Colab

Detects sponsor logos in F1 broadcast footage with YOLO, as a first step towards measuring each brand's screen time. Zero-shot YOLO-World finds no real logos; a YOLO11s fine-tuned on 69 labelled frames reaches mAP50 0.61 on held-out frames.

Read the write-up →

2025 – present

LangChain, Ollama, Phi-3, Qdrant, Sentence Transformers, MLflow, Streamlit

A local assistant that answers programming questions from a code corpus, reviews its own answers and retries when they fall short.

Bar chart of correctness and faithfulness per agent. The Phi-3 self-correcting baseline scores 83% correctness; tool-calling agents on 3B models score between 8% and 33%.Bar chart of correctness and faithfulness per agent. The Phi-3 self-correcting baseline scores 83% correctness; tool-calling agents on 3B models score between 8% and 33%.
LLM-as-judge benchmark. The self-correcting pipeline reaches 83% correctness; the best 3B tool-calling agent reaches 33%, as its loops often hit the step limit before answering.

Read the write-up →

2026 – present

PyTorch, tiktoken, PPO, RLHF

A Llama-style language model written from scratch in PyTorch, small enough to pre-train on a laptop, then fine-tuned with reinforcement learning.

Line chart of training and validation cross-entropy loss over 2,000 steps, falling from 6.0 to about 3.3.Line chart of training and validation cross-entropy loss over 2,000 steps, falling from 6.0 to about 3.3.
Pre-training on TinyStories. Cross-entropy falls from 6.0 to about 3.3 over 2,000 steps, with validation loss tracking training loss closely.

Read the write-up →

2025

Part of my MSc thesis

PyTorch, D4RL, MuJoCo, Guild AI, Hugging Face Hub

A reimplementation of the MOReL offline reinforcement learning framework with the improvements I proposed in my MSc thesis. It outperforms the original on the D4RL HalfCheetah benchmark.

Line chart of average return over 2,500 policy-search steps. J-MOReL levels off around 6,700; the original MOReL levels off around 5,950.Line chart of average return over 2,500 policy-search steps. J-MOReL levels off around 6,700; the original MOReL levels off around 5,950.
Average return during policy search on D4RL halfcheetah-medium. J-MOReL levels off around 6,700, against about 5,950 for the original MOReL.

Read the write-up →

Experience

AI Engineer

MediaRadar

Nov 2026 – present

Madrid, Spain

Python, Kubernetes, Argo CD, PostgreSQL, Databricks, GCP, Azure DevOps

  • Building machine learning pipelines that classify media content, deployed on Kubernetes.

AI Solutions Architect

Multiverse Computing

Feb 2026 – Jun 2026

Zaragoza, Spain

LLMs, RAG, Agents, Fine-tuning, Whisper, Computer vision

  • Met clients in health, climate, media and manufacturing to understand their problems, then designed LLM-based solutions: model compression, fine-tuning, retrieval-augmented generation and agent orchestration, plus computer vision and Whisper speech pipelines.
  • Wrote the technical side of proposals for private companies and of public R&D consortium bids, including EU Horizon and Spain’s Misiones programme.
  • Benchmarked models and built demos to validate proposed solutions.

Research Software Engineer

Aragón Health Research Institute (IIS Aragón)

Oct 2020 – Jul 2021

Zaragoza, Spain

C#, Kotlin, Android, Deep learning, Edge deployment

  • Built core components in C# and Kotlin for TrackAI, a platform that uses eye tracking to assess visual function in patients who cannot take a standard eye test, such as infants. It has been validated with more than 4,000 patients.
  • Deployed deep-learning gaze-estimation models to mobile hardware (Huawei NPU) through an Android app.
  • Worked with clinicians and researchers to replace manual evaluation steps with automated diagnostic tools.

Research Intern

University of Zaragoza, Graphics & Imaging Lab

Jul 2020 – Sep 2020

Zaragoza, Spain

Python, Data analysis, Visualisation

  • Studied how visual perception develops during infancy through eye-tracking experiments.
  • Processed high-frequency gaze recordings into visualisations the research group could use for large-scale analysis.

Machine Learning Intern

Hiberus Tecnología

Jul 2019 – Aug 2019

Zaragoza, Spain

Python, XGBoost, Neural networks, JavaScript

  • Built binary classifiers (XGBoost, neural networks) that predict customer conversion from real sales data, with feature-importance analysis, and presented the results to the R&D team.
  • Developed blockchain web applications on Ethereum smart contracts.

Education

MSc in Computer Science

University of Innsbruck

Oct 2021 – Sep 2025

Innsbruck, Austria

Specialisation in Robotics, Perception and Interaction. Final grade: Excellent, 96/100.

BEng in Computer Science Engineering

University of Zaragoza

Sep 2016 – Oct 2020

Zaragoza, Spain

Specialisation in Computing.

Research

RL-DOVS: Reinforcement Learning for Autonomous Robot Navigation in Dynamic Environments

Journal article, 2022

A. K. Mackay, L. Riazuelo, L. Montano. Sensors 22(10), 3847, 2022.

A reinforcement learning planner for service robots that navigate among people moving unpredictably. Instead of learning from raw lidar or camera data, the agent sees the scene as a dynamic object velocity space (DOVS): which velocities are safe for the robot, given its kinodynamic constraints and the obstacles’ motion. This representation makes training faster than learning from raw sensors.

BibTeX
@article{mackay2022rldovs,
  title   = {{RL-DOVS}: Reinforcement Learning for Autonomous Robot Navigation in Dynamic Environments},
  author  = {Mackay, Andrew K. and Riazuelo, Luis and Montano, Luis},
  journal = {Sensors},
  volume  = {22},
  number  = {10},
  pages   = {3847},
  year    = {2022},
  doi     = {10.3390/s22103847}
}

Model-Based Reinforcement Learning: Optimizing Action Choice on Learned Dynamics

MSc thesis, 2025

MSc thesis, University of Innsbruck. Awarded the highest grade.

Learns a model of the environment’s dynamics from offline data with neural-network ensembles, then uses it to choose actions while penalising predictions the ensemble disagrees on. The work is public as three projects: J-MOReL, an improved reimplementation of the MOReL offline RL framework; MBiCEM, a trajectory planner that runs iCEM on learned models; and the dynamics ensembles both are built on.

Reinforcement Learning Applied to Autonomous Multi-Robot Navigation Systems in Dynamic Environments

BEng thesis, 2020

BEng thesis, University of Zaragoza. Extended into the RL-DOVS journal article.

Skills and languages

ML & LLMs
PyTorch, Hugging Face Transformers, LangChain, scikit-learn, XGBoost, Ultralytics YOLO, OpenCV, Reinforcement learning
Data
Databricks, PySpark, SQL, Pandas, Qdrant
MLOps & cloud
Docker, MLflow, FastAPI, GCP (Cloud Run, IAM, Artifact Registry, Secret Manager), CI/CD, Git, Linux
Languages
Python, C++, C#, Kotlin, Java, JavaScript
Visualisation
Streamlit, Tableau, Matplotlib, Seaborn
Spoken
Spanish (native), English (fluent, Cambridge C2 Proficiency), German (A1)

Contact

Working as an AI Engineer at MediaRadar. The quickest way to reach me is by email at [email protected]. My code is on GitHub, and I am also on LinkedIn.