Software Engineer, Applied AI
Remote; London, UK; Cambridge, UK · 3 weeks ago
Build LLM agents that inspect circuit boards, program robots, monitor the factory and automate back-office work, with evals proving each agent works before it runs alone. Strong software engineering is central, including concurrency, idempotency and security.
- Workplace
- Remote
- Salary
- 80K GBP/yr - 250K GBP/yr
- Team
- Software Engineering
Skills
About the role
About CircuitHub CircuitHub is reshaping electronics manufacturing with The Grid, a factory-scale robotics platform that makes small-batch, high-mix electronics assembly radically more efficient. Think semiconductor-fab precision applied to the chaotic world of prototyping and low-volume production: a 10x throughput improvement in one of the world's most foundational industries. We've raised $45M from top-tier investors including Plural, Y Combinator and Google Ventures, and we're already profitable. Our customers include a major EV maker, a top AI lab and many of the worlds largest companies. The Grid isn't a prototype. It's live, scaling fast and delivering real revenue, and we're building the team that will scale it further. The Role CircuitHub makes electronics on demand. Engineers upload a circuit board design; we buy the parts, build the board in our own robotic factory and ship it. More and more of that business is now run by AI agents, and we want you to help build them. These aren't chatbots. Our agents have real jobs: • They inspect boards. An agent checks camera images of every panel we build for wrong, rotated or missing parts. • They program robots. Agents read datasheets so our pick-and-place machines know how to handle each new part. • They watch the factory. They spot stray parts in robot storage, diagnose printer faults and clear stuck machines. • They run the back office. They book supplier invoices, chase late orders, explain quotes and triage bugs. • They write our code. Most of what our team ships starts in an agent session. What you'll do Find a process people still do by hand and automate it end to end: decide what needs judgement and what needs plain code, build the tools, and prove it works before it runs alone. Then make it cheaper, faster and right more often. You'll also improve the platform every agent runs on, and work closely with the people your work serves, from factory operators to finance. How we build • Evals first. An agent earns the right to work alone: blind tests against people's verdicts, shadow runs beside the humans, replays of past decisions. We track accuracy and cost per run, and re-test whenever we change a model, a prompt or the effort level. • Agents judge, code writes. Agents return structured answers. Plain code checks them against a second source and makes every change. • Classical first, model second. Cheap vision and maths narrow the problem so the model only looks where it matters. • Safe by construction. Agents work in sandboxed cloud machines that never hold real credentials, and tool calls are built so a business action can't run twice. Recent work • Image alignment ranks the parts on each board photo. The worst 20% of parts hold about 90% of real defects, so the agent only inspects those. In a blind re-run it caught all 13 defects operators had marked. • A photo check for stray parts in robot storage scored 10 out of 10 blind, at about 8¢ a photo. The model it replaced caught about half. • Agent tool calls leave durable receipts, so a session killed mid-deploy resumes without sending the same email or order twice. • Invoices extracted from supplier emails matched what a person had entered 66 times out of 66. • A constraint solver replans the whole factory every two minutes, across robots, ovens and people. What we look for • You've put LLM agents into production and can tell us what broke and how you knew it was working. • You're a strong software engineer first. Most of this job is real engineering, not prompting: concurrency, idempotency, data modelling, security. • You build the eval before you trust the result. • You like the physical world: cameras, robots, parts rotated by 90°. • Bonus: computer vision, optimisation or electronics. Stack • Agents: Claude Code, Codex, Claude and OpenAI APIs, MCP, skills in plain Markdown • Code: TypeScript (Bun, React), Python (FastAPI, Dagster, OpenCV, OR-Tools), PostgreSQL • Infra: Nix and NixOS, Nomad, Terraform, Vault, AWS, Modal sandboxes, OpenTelemetry Pay UK £80k–£125k · US $125k–$250k, depending on experience, plus equity.
Posted Oct 18, 2026