Sofiia Druchyna

I am a fully-funded second year Master of Science & Engineering (MSE) student at Princeton University’s Computer Science Department, advised by Professor Elad Hazan.

I received a Bachelor of Science (BSc) degree (Summa Cum Laude) with Honors in Computer Science and Cybersecurity from the Tulane University, New Orleans. Upon my graduation I was inducted into the Phi Betta Kappa and Chi Alpha Sigma Honors Societies.

My research experience spans machine learning theory, online control and optimization, DL/RL, computer architecture, computer systems, LLMs in education, and cybersecurity policy. My current research interests are in AI and Education, Algorithmic Decision-Making, HCI, Causal Inference, Online Learning and Optimization.

Email | CV | LinkedIn | ResearchGate

Projects

Language Models as Teaching Assistant Companions: Evidence from Experiments in a Proof-Based Course (Submitted to CHI 2026)
Romina Mahinpei, Sofiia Druchyna

Empirically demonstrated how and when LLMs can support teaching assistants in proof-based courses through a multi-part case study in a real undergraduate class—systematically comparing LLM and human grading decisions across expertise levels and evaluating TA preferences for LLM-generated feedback to motivate human–AI designs that separate evaluative judgment from formative support.

CNPE: A Human-Centered Framework for Integrating Interactive Theorem Provers into Proof Education (SIGCSE TS 2026) [ACM Library]
Romina Mahinpei, Sofiia Druchyna, Xinran Bi

Poster presented at iRAISE Workshop as a part of AAAI 2025 Conference in Philadelphia, PA.
Explored the integration of interactive theorem provers (ITPs) into undergraduate proof-based Computer Science (CS) courses at Princeton University, emphasizing the importance of addressing stakeholder challenges and needs before implementing AI solutions. We highlight both the potential and limitations of ITPs, advocating for a human-centered approach to responsibly integrate AI tools into education, which can serve as a model for diverse educational contexts.

Efficient Spectral Control of Partially Observed Linear Dynamical Systems (NEURIPS 2025) [arXiv]
Anand Brahmbhatt*, Gon Buzaglo*, Sofiia Druchyna*, Elad Hazan

Proposed a new method for controlling linear dynamical systems under partial observation and adversarial disturbances. Our new algorithm, Double Spectral Control (DSC), matches the best known regret guarantees while exponentially improving runtime complexity over previous approaches in its dependence on the system’s stability margin.

A New Approach to Controlling Linear Dynamical Systems (ICLR 2026) [arXiv]
Anand Brahmbhatt*, Gon Buzaglo*, Sofiia Druchyna*, Elad Hazan

Poster presented at NEURIPS 2025 DynaFront (San Diego, CA) and NYRL 2025 (New York, NY) Workshops.
Proposed a new method for controlling linear dynamical systems under adversarial disturbances and cost functions. Our algorithm achieves a running time that scales polylogarithmically with the inverse of the stability margin, improving upon prior methods with polynomial dependence maintaining the same regret guarantees.

SpectraLDS: Distilling Spectral Filters into Constant-Time Recurrent Models (NEURIPS 2025) [arXiv]
Devan Shah, Shlomo Fortgang, Sofiia Druchyna, Elad Hazan

Introduced the first method, SpectraLDS architecture, for system identification of a symmetric linear dynamical system with provable robustness guarantees and with performance independent of the effective memory of the system or state dimension of the underlying system.

Honors Thesis: “Enhancing Teaching Methods of File Systems from the Educational Perspective” [Paper]

Developed an innovative teaching approach that introduces undergraduate students to the concepts of file systems through hands-on experience. WaveFS is a new lab assignment for the Computer Systems and Networking class at Tulane University, focusing on creating a simple Unix-like file system using the File System in Userspace (FUSE) interface. The lab assignment involves implementing basic file system operations to read and write data, as well as manage directories and files.

Predictive Models for Cyber Threat Trends Analysis [Code] [Paper]

Provided insights into cyber threat actors’ behavior and motives using machine learning methods, particularly KNN classifiers as well as identified common cyberattack trends using advanced data science techniques. The findings can be used for predictive analysis of future cyber threats and to enhance global threat intelligence.

MIPS R4000 Simulator for an 8-Stage Pipeline [Code]

Built a MIPS R4000 simulator (from the MIPS64 family) for an eight-stage pipeline. The processor uses deeper pipeline than the simple 5-stage design. This deeper pipeline allows it to achieve higher clock rates by decomposing the five-stage integer pipeline into 8 stages. Since the cache access stages are usually on the critical path, the extra pipeline stages come from decomposing the memory access stages to achieve more balanced pipeline stages.

RISC-V Disassembler [Code]

Built a disassembler that takes in a binary file (or a text file) as input and displays the equivalent RISC-V assembly instructions along with the binary code.

Publications

Romina Mahinpei, Sofiia Druchyna, and Xinran Bi. February 2026. CNPE: A Human-Centered Framework of Challenges and Needs in Proof Education. SIGCSE Technical Symposium 2026. ACM Library.
Anand Brahmbhatt*, Gon Buzaglo*, and Sofiia Druchyna*. January 2026. A New Approach to Controlling Linear Dynamical Systems. ICLR 2026, NYRL Workshop. arXiV.
Anand Brahmbhatt*, Gon Buzaglo*, and Sofiia Druchyna*. September 2025. Efficient Spectral Control of Partially Observed Linear Dynamical Systems. NEURIPS 2025, DynaFront Workshop. arXiV.
Devan Shah, Shlomo Fortgang, and Sofiia Druchyna. September 2025. SpectraLDS: Distilling Spectral Filters into Constant-Time Recurrent Models. NEURIPS 2025. arXiV.
Craig P. Orgeron, William Rials, and Sofiia Druchyna. January 2025. The AI-Driven State: How Government-as-a-Service Is Transforming Public Service. International Journal of Electronic Government Research (IJEGR), 21(1), 1–24. IGI Global Scientific Publishing.

* Authors contributed equally to this work.

Professional Experience

Software Development Engineer Intern, Amazon (Summer 2025)

AGI High Performance Computing Team.
Built HPC and ML infrastructure features that cut testing cycles by 14 days and enabled seamless onboarding of 1,000+ concurrent training workflows with 80% greater scalability and efficiency.

Program Analyst Intern (AppSec), BeyondTrust (Spring 2024)

Improving the SAST operations for code scanning by analyzing the existing coverage gaps and researching alternative solutions to create effective operational processes and introduce more robust scanning methods.

IT Department Assistant, Tulane University (2022-2024)

Organizing and Executing Cybersecurity Competition Events.
Managing Information Technology Departmental Operations.

Teaching

I have had the privilege of contributing to the instruction of the following courses:

COS 324 Introduction to ML, Princeton University (Assistant Instructor/Head TA) (Spring 2025 - Spring 2026)
COS 316 Computer Systems Design, Princeton University (Assistant Instructor) (Fall 2024)
CMPS 2300 Computer Systems and Networking, Tulane University (Assistant Instructor) (2023-2024)
CPMPS 1600 Introduction to Computer Science II, Tulane University (Teaching Assistant) (Spring 2023)