Principal AI Engineer

About LEC AI LEC AI is an applied AI company within LEC Group, whose parent — London Export Corporation — has been trading internationally since 1951. We build Donnie, the organisational intelligence platform: always-on organisational memory that unifies voice, email, calendar, documents, messages, CRM and meetings into one live context an entire department can query. Donnie entered production in 2026, and Bishop, our data intelligence layer, followed. We are a small team in London shipping to real enterprises, which means the work you do reaches people who depend on it quickly. The Mission We are building an organisation that runs on agents — not a single department, but the full structure. Every business function staffed by teams of specialist AI agents: analysts, marketers, financial minds, editors, operators and researchers, organised into departments with departments beneath them, working against living knowledge bases continuously fed from our own companies' data and from knowledge harvested from the outside world. They coordinate in structured swarms, debate and verify each other's conclusions, propose plans that humans approve, and learn permanently from outcomes. Agents that create, test and improve other agents under proper governance — and agents that write, run and rework their own code inside sandboxed, isolated environments, so the system's capability grows continuously and safely. The world's frontier has demonstrated what coordinated agent swarms can do at scale. Our ambition is the commercially disciplined version: swarms that run businesses. You will be the person who builds this — the architecture, the coordination, the memory, the learning loops, and the governance that keeps it safe and auditable. Responsibilities Swarm-scale architecture- Design and build many specialist agents in structured squads and hierarchies — a full organisational structure of agent departments — coordinating over durable infrastructure with defined roles, handoffs, reconciliation and containment.Sandboxed self-coding- Build environments where agents safely write, execute, test and rework their own code — fully isolated, permission-gated, auditable, with every change versioned and reversible.Expert agent modelling- Turn the recorded thinking of genuine specialists — documents, transcripts, recordings, decisions — into agents whose knowledge lives in files the system reads, extends and rewrites as understanding improves, continuously refreshed from the field.Self-extension under governance- Design mechanisms by which agents create, evaluate and improve other agents and their own instructions, with every change gated, versioned and reversible.Judgement systems- Build panels of agents that assess work independently, debate with evidence, and produce calibrated verdicts that become training data.Operational spine- Establish cost governance, quota management, and human approval gates at every boundary where the system touches the outside world. Requirements You have designed and shipped production multi-agent systems — real users or real operations, not demos or notebooks — and can explain their architecture, handoffs and failure modes from experience.You have worked at swarm scale or close to it: systems of many coordinating agents, or deep experience in multi-agent research and its engineering combined with real shipping ability.You have built, or can demonstrably design, self-extending systems — agents that create and improve agents under governance. This is the strongest signal we can see.You have turned real-world expertise into working agents — knowledge modelling, retrieval, memory architectures, and learning loops where corrections genuinely stick.Reliability engineering is fundamental to how you work: retries, idempotency, queues, reconciliation, containment — you know how these systems fail because yours have failed and you fixed them.Fluency with the current open-source agent ecosystem. Hands-on experience with the leading Chinese open-source models and frameworks is a strong advantage, as is professional language fluency.You have built or worked extensively with sandboxed code-execution environments for AI systems — isolation, permissions, resource limits, rollback — and treat containment as an architectural requirement, not an afterthought.You ship production code daily. You would rather build the system yourself than direct others to build it. What We Offer We are looking for people at the top of this field — researchers who build, or builders who read the research. The one thing we offer that neither big-company labs nor academia can: your agents will run real businesses, and real revenue will be the fitness function. Full ownership of the architecture. A direct line to the Group CEO. Our own infrastructure. And a group of operating companies as the laboratory.