Kutay Buyruk

Education

Bilkent University, Turkey (2021 – 2026)

• Bachelor of Science, Computer Science
• GPA: 3.56 (High Honours)

Experience

Developer — Google (08/2025 – Present)

• Collaborating with Google Anonymization Team on Google’s open-source Differential Privacy library.
• Implementing Apache Beam pipelines with privacy identifier-based contribution bounding, reducing data loss during differentially private data generation using Go.

Software Engineer — Olympos (06/2025 – Present)

• US-based company providing data-based services and specialized software to VC firms.
• Built machine learning models to predict potential future founders using multi-source data aggregation, reducing manual inspection by 95%.
• Designed and implemented LLM-based classification pipelines for automated company relevance filtering.

Full-Stack Engineer — TÜBİTAK (06/2024 – 06/2025)

• Worked on a microservices-based system for nationwide accounting transactions.
• Developed Java/Spring Boot microservices and REST APIs; handled national-scale accounting data in PostgreSQL.

Competitive Programming Camp — Inzva (06/2024)

• Learned and practiced algorithms and data structures in C++ at Inzva’s 2023 & 2024 summer camps.
• Secured 3rd place in the final-day contest with my foundation-league team in 2023.

Projects

Inference Time Scaling Laws for Faithfulness — Neel Nanda’s MATS Submission (03/2025)

• Conducted a 16-hour research sprint as submission for Neel Nanda’s MATS stream.
• As a follow-up to Chua and Evans, investigated evidence for an inference-time scaling and CoT faithfulness relationship.
• Ran experiments on GPQA-Diamond with and without bias cues; analyzed faithfulness in normal vs. extended CoT using hard-coded and LLM judges.

Tiny Transformer Learns Modular Reversal — Published Analysis (02/2025)

• Trained a single-layer transformer to learn circuits for reverse modular addition ($\text{reverse}(X) + 1 \text{ mod } 4$).
• Analyzed attention focusing on reverse-positioned tokens and the MLP mapping $i → i + 1 \text{ mod } 4$.
• Performed ablations on attention and MLP layers to verify their function in the circuit.

Encouraging Chain-of-Thought with SAEs — Goodfire & Apart Hackathon (11/2024)

• Attempted to improve math performance on Llama-70B + SAE with longer CoT by identifying relevant concepts and tweaking SAE probes.
• Designed and ran math evaluations to determine performance changes.

MuseBot, Song Recommender AI — Personal Project (2023)

• Built an AI-powered music recommendation system with Python.
• Integrated OpenAI LLMs and embeddings into Streamlit via LangChain; implemented a RAG pipeline over a vector database of songs.

Skills

Languages: (Competent) Java, C++, Python | (Intermediate) TypeScript, JavaScript, Go
Technologies: NumPy, PyTorch, LangChain | React.js, Node.js, Spring, PostgreSQL

Organizations

Bilkent AI Society — Active Member (2023 – Present)

• Participated in a joint meeting and Q&A with OpenAI researchers on AI Safety, led by the European Network for AI Safety.

Young Entrepreneurs Society, Bilkent — Organization Division Member (2022)

• Conducted interviews with women entrepreneurs from various industries; shared insights with the community.