Computer Science Reliability Engineer

Ruotian Wu

E

+1-226-***-**** Home Page *******.*****@*****.***

DUCATION

• Master of Mathematics, Computer Science GPA:4.00/4.00 Supervisor: Pascal Poupart Sept.2024 - Sept.2026

• Bachelor of Mathematics, Honours Computer Science GPA:3.97/4.00 Sept.2019 - April.2024

• Bachelor of Mathematics, Honours Statistics GPA:3.97/4.00 Sept.2019 - April.2024 University of Waterloo Waterloo, ON, Canada

PROFESSIONAL EXPERIENCE

• Research: Reward Guided Text Generation (RGTG) - LLM Alignment Sept.2024 - Sept.2025

Benchmarked state-of-the-art RLHF approaches (e.g. PPO, DPO, RAD, ARGS, CD and VAS), and identified key limitations including inflexible training paradigms and sub-optimal heuristics.

Developed PARGS, one of the earliest RGTG methods leveraging partial-sequence preference data, enabling flexible RLHF that bypasses training expensive base models while improving generation scores by 51%. [Publication: A Critical Look At Tokenwise Reward-Guided Text Generation (Under Review COLM2025)]

Proposed FaRMA - an efficient, and principled reward model training paradigm - reducing the inference overhead of RGTG methods by a factor of 6x and further improving the generation scores by 24%.

[Publication: Towards Cost-Effective Reward Guided Text Generation (Under Review ICML2025)]

Conducted extensive experiments by reproducing all baselines and proposed methods across four datasets, leveraging Vector Institute and Compute Canada HPC resources. Performed multi-GPU and distributed training using optimization techniques such as data/model parallelization, quantization, model pruning, distillation and efficient attention, improving training efficiency by a factor of 10x.

• Scribendi Inc. Student Researcher Sept.2024 - Sept.2025

Proposed a prompt-based classifier and leveraged LLMs as Grammar Error Correction (GEC) evaluation tools. Developed evaluation tools based on reward models and OpenAI API’s, increasing human evaluation alignment scores from 40% to 92%.

Fine-tuned LLM GEC models on the Scribendi dataset (35M source-target pairs) using both local and Google Cloud Platform (GCP) GPUs. Adopted parallel training to reduce training time by 75%, increasing accuracy by 43% compared to human expert responses.

Developing a domain-specific Neuroscience Manuscript Enhancement (NME) model aligned with top-tier publishing trends, leading to an in-production tool (Copilot for Neuroscience) that helps improve the clarity, quality, and impact of manuscripts.

• AI for Fairness Research Leader Feb.2023 - Aug.2023

Designed and implemented skill-based fairness metrics to evaluate individual players and teams in online FPS games, providing quantitative backing for matchmaking decisions.

Designed an adaptive agglomerative clustering algorithm, leading to more stable matchmaking results and reducing fairness imbalance by 58%.[Publication: Achieving fairness in team-based FPS games: A skill-based matchmaking solution (MLA2023)]

• Manulife/JohnHancock Platform Reliability Engineer Jan.2023 - April.2023

Built and maintained real-time APM dashboards using Dynatrace, New Relic, and Moogsoft, improving system observability and reducing incident response time by 30%.

Automated alert reporting pipelines with custom scripts and API integrations using Python and REST, reducing manual intervention by 50%.

SUMMARY OF QUALIFICATIONS:

• AI Engineering: Pytorch, Slurm, HPC management, Distributed Training, DeepSpeed, ZeRO, TRL

• Programming Languages: Python, R, Bash, C++, C, SQL, Powershell, JavaScript

• Skills: Git, Latex, Markdown, CI/CD, REST, Postman, FineReport, Dynatrace, New Relic, MS Office

• Certificate of Computer and Software Technology Proficiency (Software designer) ADDITIONAL PUBLICATIONS

1. Uncertainty-Guided Likelihood-Tree Search, Julia Grosse, Ruotian Wu, Ahmad Rashid, Philipp Hennig, Pascal Poupart, Agustinus Kristiadi [Under Review ICML2025]

Contact this candidate