Staff ai engineer agent and inference systems

Staff AI Engineer, Agent and Inference Systems English and Mandarin speaking. Both required. LEC AI · London · Full time, in person Principal for the right person. We are looking for one person, and we are looking for the best we can find. Someone genuinely exceptional at building agent systems and running models in production, who wants a team around them and the hardest problem they will get to work on. We do not care how old you are or where you worked. We care what you have built. WHO WE ARE London Export Corporation was among the first British companies to trade with modern China. Seven decades later the group spans robotics, drinks distribution and China trade, with real revenue behind all of it. LEC AI is the jewel in the group and it is what will drive everything else. It is a scale up in every way that matters to you. Small team, early, everything to build. What it is not is precarious. There is no investor clock, no runway to worry about, and a seventy year old group with real revenue behind it. You get the pace of a startup without betting your next two years on someone else's funding round. THE TWO PROJECTS A content and marketing platform, in active build and one of our biggest focuses. Clipping, captioning, reframing, generation and scheduling across multiple brands and channels, driven programmatically rather than by hand. There is a lot still to do on it and it gets better every week. Underneath it sits a self hosted stack we own end to end: persistent memory across vector and graph stores, an audio pipeline, agent orchestration, coding agents dispatched to build things, model agnostic serving across our own models and hosted ones. And the project above it, which is why this role exists. A system that continuously watches what is emerging in AI, works out what is worth having, gets it running on our own hardware at a cost we can measure, and puts it to work. Many agents running at once, handing work between each other, some of them writing and testing code. Tens of thousands of items a day through the media pipelines. It is designed to run without a person approving each step, which means as much of the interesting engineering sits in the controls as in the capability. What an agent may do on its own and what it may never do. How you make an action reversible so that autonomy is safe rather than reckless. How agents check each other rather than being checked by us. How the whole thing is measured on outcomes rather than on activity. The agents are also expected to improve their own working methods over time, tested against held out sets rather than trusted. Nobody has built this shape of system and there is no reference implementation to copy from. WHAT IT TAKES Agents that work off each other, at scale, without losing work. Not prompt engineering, not calling an API in a loop. If you have built this you know it is a distributed systems problem wearing an AI costume: idempotency because everything retries, leases because workers die, backpressure, reconciliation, and traces you can still follow at three in the morning. Media pipelines at volume, at a cost you can defend. Speech, speaker separation, video, on screen text, document parsing. Self hosted serving, quantisation, batching, GPU utilisation. Cost per unit is a live metric here, not a retrospective. Turning research into production systems. Most of what is worth using arrives half working. Getting it reliable, fast and cheap enough to depend on is where the real work sits, and it is the part most people underestimate. Sandboxing and isolation done properly. Agents write and execute code here. That is real engineering, not a checkbox. WHAT YOU SHOULD HAVE DONE Not years served. Finished things. • Built an agent system that ran continuously and did real work, and can say honestly how it failed • Run self hosted inference in production, not just called APIs • Shipped a speech or vision pipeline at volume and know where it breaks • Built something distributed that does not lose work under failure • Taken research code to production reliability • Measured and reduced cost per unit on something real • Thought hard about sandboxing code you did not write Nobody has all seven. If you have four and the rest do not scare you, apply. THE HONEST PART This is not a balanced place and we will not pretend otherwise. It moves fast, decisions are made in hours, and getting this standing up will take an obsessive stretch of effort. If you want a calm environment with settled processes, we are not it. If you want the thing you cannot stop thinking about, this is that. WHAT YOU GET A team hired around you, with you choosing who joins it. Direct access to the Group CEO and no committees. Competitive compensation. Something nobody has built, with no playbook to follow. HOW WE ASSESS A long conversation about something hard you built, where the failures interest us more than the launches. A short paid work sample. Then an argument, where we hand you an open engineering question and you tell us why our answer is wrong. TO APPLY Email talent@lecai.ai with "Staff AI Engineer, Agent and Inference Systems" as the subject line. Send the thing you are proudest of building and a paragraph on why it was hard. A CV is optional. If you have an agent system, an inference or benchmark writeup, or a repository where you took something rough and made it work, lead with that.