Jonathan Muwanguzi

Built reinforcement learning agents for Ms. Pac-Man using the OpenAI Gym environment. Started with a baseline Deep Q-Network (DQN) and extended it with Double DQN, Upper Confidence Bound (UCB) exploration, and a dueling architecture. Systematically tuned hyperparameters and replay buffer strategies to stabilise training. The best-performing agent, the agent with UCB plus the dueling architecture achieved the highest episodic scores and outperformed the baseline vanilla DQN baseline. Beyond raw scores the different techniques produced agents with different behaviours, by studying these differences I was able to tune the agents in order to improve performance.

Designed and deployed a serverless AI pipeline on AWS to automate visual inspection of manufacturing parts. Trained a convolutional neural network in Python to detect surface defects, containerised the model with Docker, and deployed on AWS Lambda with Step Functions for orchestration. Results were logged in DynamoDB and S3 with annotated images, and system performance monitored through CloudWatch. Achieved over 70% classification accuracy and validated deployment on industrial Mini-PC hardware.