Narek Maloyan

AI Research Engineer & PhD Candidate

GitHub · LinkedIn · Kaggle · Google Scholar · Email · X/Twitter · Resume PDF

I am an AI Research Engineer at Zencoder, where I build AI-powered coding agents. In May 2025, our team reached #1 on SWE-bench Verified with a 70% success rate, setting a new benchmark for autonomous software engineering. My day-to-day work sits at the intersection of large language models and developer tools — figuring out how to make AI agents that write reliable, production-quality code.

I am also a PhD candidate at Lomonosov Moscow State University, where my research focuses on AI safety and LLM security. I study prompt injection attacks across coding assistants, evaluation systems, and tool-integrated agents — essentially, how LLMs can be manipulated and how we can defend against it. This includes work on LLM-as-a-Judge vulnerabilities, MCP protocol security, and trojan detection in large language models. To date, I have co-authored 13 peer-reviewed publications spanning AI safety, LLM security, medical AI, and computer vision.

Before focusing on AI safety, I worked as an ML engineer across several domains: video highlights and recommendation systems at Viasat, medical article recommendations and speech-to-text at TrendMD, and MRI-based brain tumor classification at Burdenko Neurosurgery Institute. I have been teaching a graduate-level Deep Learning course at MSU since 2021, and I maintain open-source projects including manim-js (339+ stars), a TypeScript port of 3Blue1Brown's animation engine.

Read the full bio →

Experience

Zencoder — AI Research Engineer (2024–present) Building AI-powered coding agents that autonomously resolve software engineering tasks. Part of the team that reached #1 on SWE-bench Verified with a 70% success rate (May 2025), using a multi-agent ensemble architecture with parallel execution and critic-based solution selection. Working with LLMs including Claude and GPT models, agent orchestration frameworks, and evaluation infrastructure for benchmarking coding agent performance.

Viasat — ML Engineer (2021–2024) Developed ML systems for a major video streaming platform, including automatic video highlights extraction from full-length content, personalized movie recommendation engines, and semantic search optimization for content discovery. Built a highlights pipeline that automatically identified key moments in sports and entertainment content, reducing manual editing time significantly. Tech stack included Python, PyTorch, Elasticsearch, and cloud-based inference infrastructure.

TrendMD — ML Engineer (2019–2021) Built recommendation and NLP systems for a medical content platform serving millions of researchers and clinicians. Developed a medical article recommendation engine that surfaced relevant research papers based on reading context, a speech-to-text system tailored for medical terminology used by doctors during clinical documentation, and uplift modeling for optimizing content engagement. Worked with transformer-based NLP models, recommendation algorithms, and large-scale content processing pipelines.

Burdenko NSI — ML Engineer (2020–2022) Applied deep learning to clinical neuroscience problems at Russia's leading neurosurgery center. Developed non-invasive brain tumor classification models from MRI scans that could predict glioma grade without surgical biopsy, and built stroke risk prediction systems from medical imaging data. Published peer-reviewed research on glioma grading and surgical instrument segmentation, collaborating directly with neurosurgeons to validate clinical relevance. Used PyTorch, medical image processing libraries, and 3D convolutional architectures.

ODS.ai — Community Contributor (2020–2022) Data science competition pipelines, organized ML competitions with 1000+ participants.

Teaching

Watch on YouTube

Selected Publications

• Prompt Injection Attacks on Agentic Coding Assistants. IJOIT 14(2), 2026. [paper]
• Breaking the Protocol: MCP Security Analysis. Modern IT and IT-education 21(3), 2026. [paper]
• Investigating LLM-as-a-Judge Vulnerability to Prompt Injection. IJOIT 13(9), 2025. [paper]
• Adversarial Attacks on LLM-as-a-Judge Systems. arXiv:2504.18333, 2025. [paper]
• Prompt Injection Attacks in Defended Systems. DCCN, 2024. [paper]

→ Full list on Google Scholar

Blog

→ All posts

Contact

Available for full-time roles, research collaborations, consulting, and speaking engagements.

• [email protected]
• github.com/maloyan
• linkedin.com/in/nmaloyan
• x.com/NarekMaloyan
• Google Scholar
• kaggle.com/narek1110