RL environment design for coding and computer-use agents. Reward shaping, reward hacking detection, trajectory collection for browser-based tasks.
Mario Brajkovski
Building RL environments and evaluation systems for frontier AI agents
Munich · Remote · CV
Research engineer. Until recently Member of Technical Staff at HUD (YC W25, employee #7), building RL training environments and evaluation infrastructure for frontier AI labs. Co-author of ZeroDayBench (ICLR 2026 Workshop), a benchmark for whether LLM agents can autonomously discover and patch zero-day vulnerabilities. Most of my time goes into the infrastructure that makes models better: RL environments, reward design that holds up under hacking, and evals that measure what they claim to measure.
What I Work On
Evaluation infrastructure for frontier models. Figuring out what to measure and how to measure it without fooling yourself.
Agent safety and red-teaming. Trying to break autonomous agents before they go to production.
Research
Can frontier LLMs autonomously find and patch real vulnerabilities? We tested GPT-5.2, Claude Sonnet 4.5, and Grok 4.1 on 22 novel critical zero-days in open-source codebases. Short answer: not yet. But the failure modes are telling. A couple of frontier labs are now looking at running it internally.
Privacy-preserving ML using AMD SEV-SNP encrypted VMs. Full memory encryption and attestation with <20% overhead on 110M-param BERT.
Selected Work
RL Training Environments
Sixteen environments built end to end: zero-to-one coding in a blank workspace, software engineering across Go, Rust, Python, and TypeScript, full virtual desktops for computer-use agents, knowledge work with native Office deliverables, enterprise tool use, Game Boy games, robot manipulation in a physics sim, plus healthcare record systems and archived live websites for browser agents. 1,000+ tasks shipped to two of the top three frontier labs, with reward hacking analysis and prevention throughout. Several are public templates.
Evaluation Infrastructure
Eval infrastructure for an unreleased frontier model. Designed the evals from scratch, including the scoring and validation.
Verifiable Task Design
A task is only worth training on if a high score means the work was actually done. I treat every grader as something to attack: run it against deliberately broken solutions and cheating agents, check that the prompt and grader ask for the same thing, and calibrate difficulty across model tiers before anything ships.
Eval Quality & Data Supply
Automated QA that separates grader bugs from genuine agent failures, and LLM-judge failure-mode analysis over agent trajectories. Ran task quality across 5+ external vendor teams and most of the customer-facing technical communication.
Zero-to-One Coding Evals
Turning a well-tested Python or TypeScript module into an eval where an agent rebuilds it from a blank workspace: source analysis, prompts checked against the code, prompt-grader alignment checks, and hidden tests protected against cheating.
Projects
Voice-first AI companion with paying subscribers. A Live2D avatar that talks, emotes, and moves in sync with the conversation; sub-second voice latency over WebRTC, persistent memory.
AI tutoring platform for the Macedonian school curriculum, serving 320+ active students. Full-stack TypeScript with curriculum-aligned prompting and Stripe payments.
Background
2019–2025: Site reliability engineering at Joyn (streaming platform, 4 yrs), then DevOps consulting for HanseMerkur (insurance). Kafka, AWS, Kubernetes, large-scale ops.