AI Researcher & Consultant

Martin
Mwarangu

MSc Artificial Intelligence with Distinction — University of East London

I build and study AI systems that reliably do what we intend them to do. My work spans reinforcement learning alignment, value misalignment in autonomous agents, and large-scale machine learning applied to real-world problems.

View Research ↓ See Projects
MSc
with Distinction
8.7M
Records processed
IEEE
Under consideration
Martin Mwarangu at graduation
RL
Alignment
AI Safety
Research
Background

Researcher, engineer,
practitioner

I am an AI researcher and consultant based in the UK, working at the intersection of reinforcement learning, alignment research, and applied machine learning. My academic background is grounded in the MSc Artificial Intelligence programme at the University of East London, where I graduated with Distinction.

My research focuses on a problem that is often theoretical but rarely tested empirically in current systems: what happens when an autonomous agent learns to pursue a goal, and then encounters an incentive to deviate from it? I study this through gridworld experiments, reward trap analysis, and the evaluation of mitigation strategies in RL deployments.

Alongside research, I collaborate with academic partners on alignment papers and consult on applied ML projects, including large-scale classification, distributed data processing, and computer vision systems.

AI / Machine Learning
Reinforcement Learning Q-Learning XGBoost LightGBM CNNs Transfer Learning SMOTE
Data & Engineering
Apache Spark Python SQL ETL Pipelines Firebase React
Frameworks & Tools
scikit-learn TensorFlow PyTorch OpenAI Gym Git
Research

Where curiosity meets
consequence

Value Misalignment in Modern Agentic Reinforcement Learning Systems

Autonomous agents are now capable of working independently towards set goals. The alignment problem arises when these agents pursue goals contrary to their designer's intention, or ones that fail to fully capture the spirit of the task. While much alignment research is theoretical, focusing on scenarios where sophisticated AI seeks power, this dissertation investigates the problem in systems we have today — the building blocks of the systems we will have in the future.

A custom gridworld environment was designed and a Q-learning agent introduced, mirroring the classic mouse-in-a-maze experiments used to study learning and spatial navigation in mammals. After the agent demonstrated it could solve the maze, a reward trap was introduced — analogous to offering a bribe to an official — designed to encourage bad behaviour. The study then examined whether mitigation strategies could correct it.

The findings show that agents will almost always fall into the reward trap, underlining the genuine danger of misalignment. The paper argues that no single mitigation strategy is universally effective: success is situational. The conclusion advocates for holistic, robust alignment best practices and meaningful guardrails when building and deploying reinforcement learning agents.

Under consideration — IEEE Access Reinforcement Learning Value Alignment Reward Hacking Q-Learning AI Safety

Ongoing Research Interests

Beyond the dissertation, I am actively engaged in alignment research touching on agentic RL systems in production environments, the failure modes of reward shaping in real-world deployments, and the philosophical dimensions of intent specification. I collaborate with Dr Tauseef Ahmed on extending this work into broader alignment frameworks.

Agentic Systems Reward Shaping AI Ethics Deployment Safety
Applied Work

Projects that
shipped

↗

Web Domain Value Prediction

Built a machine learning pipeline to predict the commercial value of web domains, working across a dataset of 8.7 million records with a 95% class imbalance. The project required significant preprocessing, class balancing via SMOTE, and model comparison across XGBoost and LightGBM implementations to achieve reliable predictions.

8.7M records
95% class imbalance
XGB + LGBM ensemble
XGBoost LightGBM SMOTE scikit-learn Python
View Paper
⚡

Real-time Web Log Analytics

Designed and built a distributed analytics pipeline using Apache Spark to process web server logs in real time. The system used regex-based ETL to extract structured data and produced business intelligence recommendations from raw access patterns.

Apache Spark Python Regex ETL Distributed Systems
View on GitHub
🦅

Bird Species Classifier

Trained a convolutional neural network for bird species classification using transfer learning from ResNet and EfficientNet backbones. Applied standard data augmentation and fine-tuning techniques to achieve high accuracy on a multi-class dataset.

CNN ResNet EfficientNet Transfer Learning PyTorch
View Paper

Let's talk

Whether you are working on an alignment problem, an applied ML project, or simply want to discuss the research — I am always open to a conversation.

LinkedIn GitHub Twitter Email