Maze is building an AI-native vulnerability management platform. Our autonomous agents investigate, triage, and remediate security findings the way a senior analyst would, only faster and at scale. As Head of AI, you'll own the intelligence that makes those agents work: the AI research and implementation strategy for the whole company, plus the crown-jewel technical problem underneath it. Our investigation agents run multi-step, non-deterministic trajectories across a toolset of 180+ tools, tested against a ground-truth exploit lab we built for exactly this purpose. Knowing whether they're getting better, and making them better, is the most important technical problem at Maze. It's the heart of this role.
This is a hands-on leadership role, not a management layer. You'll set AI direction as a member of the engineering leadership team reporting to the CTO. But you'll spend most of your time building: designing evaluation frameworks for non-deterministic agents, running fine-tuning and model-routing experiments against real data, prototyping new techniques and getting them into the product. You'll lead a small, strong AI team (3–4 engineers today) by setting the technical bar and doing the work alongside them, while working closely with our ML tech lead and the product teams building agents day to day. Your impact comes from what you ship, not the size of your org. We're not looking for someone to run a large team from two layers up. We're looking for someone who wants to define how generative AI transforms cybersecurity and keep their hands on the code.
This role suits a deep LLM-era practitioner who has shipped agentic systems to production, can reason about transformer internals and fine-tuning from first principles, and moves fast. We're a three-product company with a lot of surface area, a well-funded Series A (Theory Ventures) behind us, and a Series B on the horizon. The AI foundation you set now becomes the moat we compete on for years — this is a foundational hire whose standards will shape Maze's AI trajectory well past this raise.
Your Contributions to Our Journey:Own AI strategy and research direction: Set the technical roadmap for our AI capabilities. Stay ahead of the research curve to find, validate, and prioritise the techniques that differentiate Maze, and turn what's real into a concrete, sequenced roadmap while discarding the hype.
Own agent quality and evaluation: Build and run the frameworks that tell us whether our investigation agents are improving. That means trajectory evaluation, ground-truth scoring against the exploit lab, and end-to-end benchmarks for non-deterministic, multi-step behaviour. This is the core problem of the role.
Build the breakthroughs yourself: Prototype a new technique in days, get it into the product, and measure the impact. You'll spend most of your time hands-on in the codebase, acting as the technical product manager who guides it to production.
Run fine-tuning and model experiments on real data: Own fine-tuning pipelines, context engineering, model migration, and cost/routing optimisation grounded in production data, not proofs of concept.
Guide prioritisation across the AI team: New techniques, papers, and ideas surface constantly. You'll be the filter deciding which of them are actually worth a prototype this week, and which are noise - keeping the team focused on what moves the needle.
Lead a small team by doing: Set technical direction for the AI engineers, raise the bar through pairing and review, hire as we scale, and stay close enough to the work to make the hard architectural calls yourself.
Partner with the CTO and engineering leadership: Turn the AI roadmap into shipped capability, and make sure evaluation is wired into how the whole team builds.
Get in front of customers: Occasional direct customer exposure, translating what security teams need into concrete improvements to the ML pipeline.
Set the pace: Ship prototypes in days, not quarters. Bring urgency to a domain where most of the field still moves slowly.
Hands-on technical leadership: A track record of leading AI work while personally building it. Strategy and implementation. You lead from the front. If you've moved permanently into management and stopped shipping, this isn't the right fit.
Shipped LLM/agentic systems to production: You've built and run generative-AI systems that real customers use, not research prototypes or slideware. You can point to agents or LLM features you put into production and improved over time.
Deep LLM-era technical depth: You can explain transformer architecture, training, fine-tuning (e.g. LoRA), and inference from first principles. We test this directly. A strong pre-LLM ML pedigree (RL, NLP, recommendations, ASR) is valuable but won't substitute for modern generative-AI depth.
Built evaluation frameworks for non-deterministic systems: You've designed and run evals for multi-step, non-deterministic agents: trajectory evaluation, LLM-as-judge, fine-tuning result measurement. This capability is rare and it's the one we most want. It will set you apart.
Top-tier pedigree with a builder's edge: Experience at a leading AI organisation or strong AI-native startup where you raised the technical bar rather than coasted on the brand.
Unambiguous startup signal: You've operated at early stage or built something from zero. You move fast, own outcomes end-to-end, and don't need a large org around you to ship. Founder experience is a strong plus.
Pace and urgency: You ship prototypes in days. You make pragmatic calls on models, cost, and scope to keep momentum, and you're impatient with quarters-long cycles.
Sharp, concise communication: You communicate clearly and tightly in a remote-first, English-speaking team, in writing and live. You get to the point.
Nice to Haves:
Security, vulnerability-management, or adversarial-domain background. Strongly preferred. Every candidate we've rated highly has had it. Offensive security, vuln management, threat detection, or applying AI to security problems all count.
Comfort in front of customers, able to translate agent behaviour and capability into terms a security team understands.
Model cost/routing pragmatism: real experience cutting inference cost and migrating between models in production.
Track record at a successful AI-first startup, scaling a system from experimentation to production impact.
PhD or published work in ML/AI at top-tier venues, paired with real production experience.
Days 1–30: Get fully up to speed on every agent we've built and how our ML evaluation pipeline works today. Start drafting a short, mid, and long-term technical plan.
Days 31–60: Ship something fundamentally new — for example, fine-tune a small model and get it into production.
Days 61–90: Move onto bigger bets — RLHF for specific use cases, scalability of the evaluation approach, and deeper customer-facing model tuning.
Manager-of-managers or 2nd/3rd-line leaders who direct rather than build.
Fractional, advisory, or part-time profiles.
Research-only backgrounds without production shipping experience.
The hardest problem in the field, unsolved. Evaluating non-deterministic, multi-step agents against ground truth is an open problem, and we've built the exploit lab and 180+ tool agent infrastructure to attack it. You'd own it, at the intersection of generative AI (LLMs and agents) and cybersecurity.
A team you'll want to be measured against. Founders and engineers from Amazon, Elastic, and Tessian. Hands-on leaders who've been part of multiple acquisitions and an IPO. Most people who join do so because of how strong the team already is.
Build the AI-native company from the ground up. A well-funded Series A (Theory Ventures) with a Series B on the horizon, early enough that you'll set the technical standards for how AI investigates security at scale.
Cybersecurity as a force for good. The work directly helps organisations stop attacks. Measurable impact, real customers, immediate feedback on what you ship.
Founding-level ownership and upside. Significant equity, a seat on engineering leadership, and a path to VP of AI as the team scales around what you build.
