Bayzhan Mukatay 🇰🇿

In my sophomore year, I identified a recurring struggle of Rice students—conflicting exam schedules.

I approached the chair of our department to mentor me in solving this problem.

Generating an exam schedule for 4,000 students was a large-scale problem, and the infeasibility of directly matching classes with classrooms made it even more exciting.

I designed a two-phase algorithm using distinct Mixed Integer Programs (MIPs): the first MIP chooses which classes to split into multiple subsections, thereby making it feasible for the second MIP to timetable the exams.

This reduced conflicts by 90% while respecting university constraints and preferences of students.

Built a web messaging app with features including authentication, workspace/channel management, and real-time threaded message interactions using TypeScript, HTML, CSS, and utilizing JSON schema validation.

Developed a Go-based network-accessible NoSQL database API, designing concurrency with Go routines for parallel execution, concurrent data structures for safe data sharing, and server-sent events (SSE) for real-time updates.

I designed a simulation-based optimization algorithm that searches for a kidney allocation policy that is equitable yet most effective in terms of post-transplant outcomes. Each policy was tested by simulating an allocation of roughly 12,000 kidneys for 90,000 waitlist candidates—a computationally heavy task. I implemented a parallel program that automatically designs and self-schedules new simulations and later gave a presentation covering our results at our center.

Implemented Monte Carlo and Q-Learning algorithms from scratch to derive optimal strategies for Blackjack, modeling it as an episodic Markov Decision Process (MDP) with discrete states and actions.

Implemented and fine-tuned a Deep Deterministic Policy Gradient (DDPG) algorithm from scratch to solve the continuous control task in OpenAI's Lunar Lander, enhancing spacecraft landing precision.

Implemented and fine-tuned a Proximal Policy Optimization (PPO) algorithm with reward shaping to solve the MiniGrid Unlock-Pickup environment, optimizing agent performance in navigation and object interaction tasks.

Developed machine learning models to predict synaptic connections in neural networks, combining computational neuroscience with modern ML techniques. Used TensorFlow to build neural network architectures that could learn patterns in neural connectivity data. This project explored the intersection of artificial intelligence and neuroscience, providing insights into how we can better understand and model biological neural systems.

The system uses object-oriented programming principles and greedy algorithms to efficiently generate comprehensive test suites that maximize code coverage while minimizing redundancy. This tool significantly improved the automated grading process for large-scale computer science courses.

Developed & simulated a new ticket allocation mechanism for FIFA World Cup using the Deferred Acceptance Algorithm with Multiple Tie-Breaking, significantly increasing first-choice ticket allocations compared to the current system.

Demonstrated the proposed model's efficiency, stability, and strategyproofness through comparative analysis and simulations.

Developed a Recurrent Neural Network (RNN) architecture with time-warping and convolutional filters using TensorFlow on Google Colab GPU to classify text sequences from three distinct datasets, achieving an accuracy of 94%.

Developed a regularized logistic regression model to classify a dataset of 170k text documents (1.9 GB) by leveraging TF-IDF method and Spark's RDDs on an AWS cluster for handling large-scale data, achieving an F1 score of 0.65.

I founded and served as President of Rice's first undergraduate INFORMS chapter, creating a thriving community of students passionate about Operations Research.

We organized numerous events with experts from industry & academia, including guest lectures, workshops, and networking sessions.

Our flagship program, DecisionLab, was a semester-long project-based initiative where over two years, 70+ students developed OR projects addressing real challenges faced by other students.

Projects included roommate matching platform, dining hall schedule optimization, staff scheduling, and many more. Along with my teammates, we created a wholesome OR community at Rice, and our efforts were recognized with Magna Cum Laude at the Student Chapter Awards ceremony at the largest INFORMS Conference in Seattle. This experience taught me the importance of building bridges between theory and practice, and how collaborative environments can accelerate learning and innovation.

Closer to the end of my sophomore year, I was reading the papers of Rice faculty in OR and found the works of Dr. Süleyman Kerimov in dynamic matching & applied probability to be fascinating. He kindly agreed to mentor me, and I spent my summer building a solid foundation by studying & reproducing the main papers and algorithms used in kidney exchange and ride-sharing platforms.

Later, Dr. Sebastian Perez-Salzar joined us for a collaboration. We worked on researching efficient ways of merging multiple dynamic matching systems, and we fully characterized the behavior of a particular dynamic matching property for merging trivial graphs. We attempted to extend this characterization towards more general graphs, which was a challenging task, but as I was starting my summer internship, this project was put on pause.

During that year, I also designed two dynamic matching algorithms (Greedy & Lazy Accumulation Policies) using the Generalized Coupon Collector Problem with Quotas. These algorithms performed well in terms of reducing waiting times and hindsight regret on simulations.

This project implements various reinforcement learning algorithms to optimize data center resource allocation and job scheduling. The system manages two types of servers (Type A and Type B) and handles both small and large jobs with different queue capacities.

I implemented and compared the performance of Value Iteration, Policy Iteration, Q-learning, Deep Q-network, and Proximal Policy Optimization algorithms.