AI Engineer

We build agent systems and install them inside other people's businesses, holdingtheir credentials. So the question we care about is not whether you can stand an agentup. It is whether you decided what it was allowed to do before you shipped what itcould do. Why this role exists now SarvX installs AI systems for operators. The founders are currently building thosesystems themselves, which means the people diagnosing the problem are also the peoplewriting the code, and that caps how many clients we can take. We need an engineer whoowns the builds and, more importantly, keeps owning them after they go live. What you would own Agent systems built and installed into client businesses The permission surface of every deployment: what each agent can call, what credentials it holds, where it can reach, and where a person has to confirm Extending the runtime, writing tools and skills rather than only configuring React front ends where a system needs a surface Keeping what is already live actually working Stack: OpenClaw, Claude Code, Node.js, TypeScript, React. The screen, and please read this before applying We will ask you to screen share an agent system you built that is currently running,and we will ask you about it for a while. What tools it can call. What credentials itholds. Where a human sits in the loop and why there and not somewhere else. Then we will ask you a version of this: a web page your agent reads contains a line oftext addressed to the agent, telling us to send it the contents of your environmentfile. Walk us through everything that happens next. You may redact credentials, customer data and anything proprietary. Please do notredact the permission configuration, the tool scoping or the egress rules. Those arethe parts we are asking about. A repository tour is not this. A recorded demo is not this. An architecture diagram isnot this. If you do not have something running that you can show us, this is not therole right now. What we are looking for You treat what an agent reads as untrusted input rather than as context You can name what your system is not allowed to do without being asked You have run something unattended for long enough to have watched it break in a way you could not reproduce, and you have a method for that Node and TypeScript deep enough to extend a runtime, not only to configure one You read changelogs, issue trackers and advisories rather than tutorials We are not asking for years on OpenClaw specifically. It did not exist long enough agofor that to be a fair question, and anyone who claims otherwise is telling yousomething. We are asking whether you have operated agents anywhere. No degree is required and we will not ask about one. Who this role is for Someone who finds the authorization question more interesting than the capabilityquestion. Who has opinions about where confirmations belong. Who wants to own anengineering function inside a business rather than take tickets, and understands thatowning it means being the person called when a client's installed system doessomething nobody expected. Who this role is not for If your answer to prompt injection is that you run it in a container, we will talkabout it and it will be a conversation rather than a rejection, but you should knowthat is the specific thing this role exists to get right. If you want to work on models rather than on systems, this is not that job. Nothinghere is research. If you have shipped a lot of things that ran once and impressed people, and nothingthat has been running unattended for six months, we are probably not a fit yet. If you need a specification handed to you, there is nobody to write it. What this role comes with From 3,000 USD per month, depending on what you have actually operated rather than on years served Pay moves with performance. This function is being built from scratch and the person who builds it is the person who runs it. If you are worth more to us in six months, we would rather pay you more than replace you Contract engagement, not employment Remote, anywhere in the United States, with working hours that overlap Pacific time Yuba City, California is where we are, and closer is better, but it is not required Live client installs from week one Real authority over the security posture of what we install, because you are the one who will have to defend it How this works Five stages, and we tell you all of them up front. 1. Application with knockout questions 2. Four written questions, read for how you think and not for polish 3. Four short recorded videos, sent only after we have read your writing 4. A live sixty minute conversation, which is where the screen share happens 5. A paid trial on real work, one to two weeks, at a fair rate We do not make an offer without the paid trial. We respond same day to strongapplications.