Narek Maloyan

AI Research Engineer & PhD Candidate

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

I build coding agents at Zencoder. In May 2025 our team hit #1 on SWE-bench Verified at 70%. Most days I'm working on the gap between "demo works" and "you'd actually let it touch your repo."

On the side, a PhD at MSU on prompt injection — how LLM agents get manipulated and how to defend them. So far that's LLM-as-a-Judge, MCP, and trojans. Thirteen papers across that, medical imaging, and CV.

Earlier: video highlights and recommendations at Viasat, medical NLP at TrendMD, brain-MRI classification at Burdenko. Teaching grad-level Deep Learning at MSU since 2021. The side project people actually use is manim-js — 3Blue1Brown's animation engine in TypeScript.

Experience

Zencoder — AI Research Engineer (2024–present) Coding agents that resolve real SWE tasks. On the team that reached #1 on SWE-bench Verified in May 2025 (70%) with a multi-agent ensemble and a critic for solution selection. Day-to-day: orchestration, evals, and shaving the long tail of failures.

Viasat — ML Engineer (2021–2024) Video highlights extraction, movie recommendations, and semantic search for a streaming platform. The highlights pipeline pulled key moments out of sports and shows so editors didn't have to. Python, PyTorch, Elasticsearch.

TrendMD — ML Engineer (2019–2021) Article recommendations and speech-to-text for a medical content network. Transformer NLP tuned for clinical vocabulary, plus uplift modeling on engagement.

Burdenko NSI — ML Engineer (2020–2022) Deep learning at Russia's main neurosurgery center: glioma grading from MRI without biopsy, stroke risk from imaging. Peer-reviewed work co-authored with neurosurgeons.

ODS.ai — Community Contributor (2020–2022) Organized ML competitions with 1000+ participants.

Projects

imast.ai

AI HR agent for sourcing, screening, and engaging talent across 800M+ profiles. Launch announcement →

AI & ML Cheatsheet

Open-source reference guide covering key concepts across ML and data science.

manim-js (339+ stars)

TypeScript port of 3Blue1Brown's Manim for creating math animations on the web.

amnezia-cloak

Minimal macOS menubar client for AmneziaWG tunnels — Swift, zero deps, ~350 lines.

polymarket-speech

Rust experiment exploring a speech-driven interface for Polymarket.

arutunyan.doctor

Practice site I built for my lovely wife, Dr. Maria Arutunyan (OB-GYN), with 26 clinical calculators and reference tools for obstetricians and gynecologists.

Teaching

Deep Learning — Moscow State University (since 2021)

Graduate-level course covering neural architectures, optimization, and practical applications.

Watch on YouTube

Selected Publications

Sleeper Channels and Provenance Gates: Persistent Prompt Injection in Always-on Autonomous AI Agents. arXiv:2605.13471, 2026. [paper]
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

The Best Fully-Local Voice Stack for Home Assistant (2026) Jun 2026

BeeLlama.cpp on a Mac Mini M4 16GB: DFlash and TurboQuant for Local LLMs May 2026

DFlash and DDTree: 8x Faster LLM Inference via Block Diffusion and Draft Trees Apr 2026

Running Qwen 3.6 Locally on a Mac Mini M4 with 16GB RAM Apr 2026

The $20K Bug That Changed How We Think About Evals Mar 2026

Contact

Open to research collaborations, consulting, and the occasional talk. Email is fastest.