{ "$schema": "https://raw.githubusercontent.com/jsonresume/resume-schema/v1.0.0/schema.json", "basics": { "name": "Mario Brajkovski", "label": "Research Engineer", "image": "https://superbavario.dev/mario.png", "email": "contact@superbavario.dev", "url": "https://superbavario.dev", "summary": "Member of Technical Staff at HUD (YC W25). Builds RL training environments and evaluation infrastructure for frontier AI labs. Co-author of ZeroDayBench (ICLR 2026 Workshop). Shipped 1,000+ computer-use and coding RL tasks to frontier labs.", "location": { "city": "Munich", "countryCode": "DE", "region": "Bavaria" }, "profiles": [ { "network": "GitHub", "username": "mariobrajkovski", "url": "https://github.com/mariobrajkovski" }, { "network": "X", "username": "super_bavario", "url": "https://x.com/super_bavario" }, { "network": "LinkedIn", "username": "mario-brajkovski", "url": "https://www.linkedin.com/in/mario-brajkovski-a12558b0/" } ] }, "work": [ { "name": "HUD", "position": "Member of Technical Staff / Research Engineer", "url": "https://hud.ai", "startDate": "2025-07", "summary": "YC W25 startup building RL environments and evaluation infrastructure for AI agents. Employee #7.", "highlights": [ "Built 16 RL and eval environments end to end: coding, software engineering, computer use, knowledge work, enterprise tool use, games, and robotics", "Shipped 1,000+ computer-use and coding RL tasks to frontier labs, with reward hacking analysis and prevention throughout", "Designed eval infrastructure for an unreleased frontier model from scratch, including scoring and validation" ], "keywords": ["RL Environments", "LLM Evaluation", "Reinforcement Learning", "Reward Hacking", "Computer-Use Agents", "Python"] }, { "name": "Cloudsurf IT Consulting", "position": "DevOps Engineer (Consultant to HanseMerkur)", "location": "Munich, Germany", "startDate": "2024", "endDate": "2025", "summary": "DevOps consulting for HanseMerkur insurance infrastructure", "highlights": [ "Led on-premises to AWS EKS migration", "Implemented Crossplane for infrastructure management" ], "keywords": ["AWS", "EKS", "Kubernetes", "Crossplane", "DevOps"] }, { "name": "Joyn", "position": "Site Reliability Engineer", "location": "Munich, Germany", "startDate": "2019", "endDate": "2023", "summary": "Site reliability engineering for streaming media platform", "highlights": [ "Managed Kafka infrastructure", "Maintained Vault for secrets management", "Built and maintained observability stack" ], "keywords": ["Kafka", "Vault", "Observability", "SRE", "Streaming"] } ], "projects": [ { "name": "ast-pilot", "description": "Python and TypeScript module to coding eval task generator", "highlights": [ "One-command pipeline: Python or TypeScript source + tests to shippable HUD coding eval task", "AST-level source scanning with LLM-powered prompt generation", "Multi-round factual validation that refuses to ship prompts drifting from code", "Prompt-grader alignment verification with auto-fix and hard-stop on contradictions", "Hidden test injection with bytecode obfuscation for anti-cheating", "Automatic repo-internal dependency remapping and import rewriting" ], "keywords": ["AST", "Python", "TypeScript", "LLM", "Evaluation", "Code Generation", "Anti-Cheating"], "url": "https://github.com/hud-evals/01-coding-template", "startDate": "2026" }, { "name": "CUA Environment Template", "description": "Open-source template for computer-use agent evaluation", "highlights": [ "Full virtual desktop stack (X11/XFCE4/Chromium) in Docker", "Dual-mode tool registration for multiple LLM platforms", "Flexible bash and LLM hybrid grading system", "Task auto-discovery with zero registration boilerplate" ], "keywords": ["Computer-Use Agents", "Docker", "X11", "Evaluation", "MCP"], "url": "https://github.com/hud-evals/cua-template", "startDate": "2026" }, { "name": "Tutor.mk", "description": "AI tutoring platform for the Macedonian curriculum", "highlights": [ "Serving 320+ active students", "Full-stack TypeScript with Next.js", "PostgreSQL database with Drizzle ORM", "Stripe payment integration", "Custom LLM prompting for curriculum-aligned responses" ], "keywords": ["Next.js", "TypeScript", "PostgreSQL", "Stripe", "AI Tutoring"], "url": "https://tutor.mk", "startDate": "2024" }, { "name": "TwoPeas", "description": "Voice-first AI companion", "highlights": [ "Sub-second end-to-end latency", "Custom WebRTC implementation with SDP negotiation and VAD", "Persistent memory system", "Real-time audio streaming", "Deployed on Firebase" ], "keywords": ["WebRTC", "Voice AI", "Real-time Systems", "Firebase", "VAD"], "url": "https://twopeas.ai", "startDate": "2024" }, { "name": "Tutorist", "description": "Real-time tutoring with Live2D avatars", "highlights": [ "Live2D avatar animation system", "OpenAI Realtime API integration", "Custom WebRTC audio session management", "Dynamic tool injection", "Emotion-mapped avatar expressions", "Infrastructure as Code with Terraform" ], "keywords": ["Live2D", "OpenAI Realtime API", "WebRTC", "Terraform", "Animation"], "url": "https://tutorist.ai", "startDate": "2024" } ], "publications": [ { "name": "ZeroDayBench: Evaluating LLM Agents on Unseen Zero-Day Vulnerabilities for Cyberdefense", "publisher": "ICLR 2026 Workshop", "releaseDate": "2026", "url": "https://arxiv.org/abs/2603.02297", "summary": "Co-author. Benchmark testing whether frontier LLM agents can autonomously find and patch 22 novel critical zero-day vulnerabilities in real open-source codebases. Evaluated GPT-5.2, Claude Sonnet 4.5, and Grok 4.1." }, { "name": "Confidential LLM inference in trusted execution environments", "publisher": "Research Paper", "releaseDate": "2024", "url": "https://superbavario.dev/thesis-1.pdf", "summary": "Demonstrated practical privacy-preserving ML using AMD SEV-SNP encrypted VMs with full memory encryption and attestation. Achieved <20% performance overhead on 110M-param BERT, showing viable path for confidential AI in production environments." } ], "skills": [ { "name": "AI Safety & Alignment", "level": "Advanced", "keywords": [ "Safety Red-teaming", "Alignment Testing", "Failure Mode Analysis", "Evaluation Infrastructure" ] }, { "name": "Machine Learning", "level": "Advanced", "keywords": [ "Reinforcement Learning", "LLM Integration", "Privacy-Preserving ML", "Trusted Execution Environments" ] }, { "name": "Real-time Systems", "level": "Advanced", "keywords": [ "WebRTC", "Voice AI", "Low-latency Architecture", "Audio Streaming", "VAD" ] }, { "name": "Full-stack Development", "level": "Advanced", "keywords": [ "TypeScript", "React", "Next.js", "Node.js", "PostgreSQL" ] }, { "name": "Infrastructure & DevOps", "level": "Advanced", "keywords": [ "AWS", "Kubernetes", "Terraform", "Kafka", "Observability" ] } ], "languages": [ { "language": "English", "fluency": "Fluent" }, { "language": "Macedonian", "fluency": "Native" }, { "language": "German", "fluency": "Professional" } ], "interests": [ { "name": "AI Safety", "keywords": ["Alignment", "Red-teaming", "Evaluation"] }, { "name": "Real-time Systems", "keywords": ["Low-latency", "WebRTC", "Voice AI"] }, { "name": "Privacy Technology", "keywords": ["Confidential Computing", "TEEs", "Secure Enclaves"] } ], "meta": { "canonical": "https://superbavario.dev/resume.json", "version": "v1.1.0", "lastModified": "2026-07-28", "note": "Machine-readable resume following JSON Resume schema. Designed for AI agents and automated systems." } }