Founding AI Engineer

About Berkeley Myles
We build software for UK manufacturers who can't afford to get things wrong: aerospace, medical devices, oil and gas, motorsport. Our product, ProgressPlus, runs the whole job from first enquiry through to final invoice. Hundreds of manufacturers run on it every day, and when one of them has to prove what went into a part, it's where the answer comes from. We're now building a new product with AI on every stage of that flow.
We're part of Banyan Software, the best permanent home for software businesses that serve specialised industries, their employees, and their customers. Banyan is a Great Place to Work Certified company, a five-time Inc. 5000 honoree, and a top-10 company on the Deloitte Technology Fast 500.
Founded in 2016 and headquartered in Atlanta, Banyan operates more than 100 portfolio companies across North America, the UK, the EU, and APAC.

About the Role
A manufacturer runs one flow: enquiry, quote, order, jobs, scheduling, tracking, quality, shipping, invoice. ProgressPlus is the system of record for that flow. We're building the system of action on top of
it: AI that works on every stage, capturing enquiries, validating orders, optimising schedules, anticipating problems, and reconciling invoices, always with a human in control and every step
auditable.  You'll be the company's first AI engineer. There's not much AI in production today: no agents, no runtime, no AI layer. You'll build the system of action from a blank page. It has two parts: the events that announce what just happened in the business, and the agents that act on them. Agents that plan multi- step work, run in the background, act under their own identity with bounded authority, and leave a trail a regulated manufacturer can show an auditor. The ground is prepared. Five .NET services with architecture boundaries enforced in CI, and a growing catalog of transaction-safe business operations ready to become AI tools. This is a founding role: you set the architecture, the standards and the evaluation discipline that every AI feature in this product gets built on.

Key Responsibilities
Build the agentic layer: one coherent layer that every AI capability in the product runs on. Once it exists, adding AI to any stage of the manufacturing flow is a few weeks of feature work.
The runtime and the tools Architect the agent runtime: a server-side loop with durable runs, retries, scheduled and event-driven triggers, and human-in-the-loop steps
Build the tool layer that turns our catalog of transaction-safe business operations into typed, tenant- scoped, permissioned AI tools, for our own agents and for the outside world through MCP
Evaluation and audit Make evals part of how we ship: fixture-based tests of every AI behaviour, wired into CI from day one Close the correction loop: every override, rejection and edit captured as structured signal, so evals and retrieval improve from real use Build agent auditing into the platform's tamper-evident action log as it lands: every agent step logged, attributable and replayable. Our customers pass AS9100 audits, and explainable AI is a
feature we sell Grounding and operations Drive the event catalog: which business facts the product announces, from the office to the shop floor (SalesOrderCreated, GoodsReceived, JobRunningLate). Every agent run starts from one Grow the semantic grounding: a machine-readable model of the business and how it fits together, so agents retrieve facts instead of inventing them
Own the agent authority model: what an agent may do, its budget, how much autonomy each workflow gets. Instrument the runtime so every step is metered and traced
Run what you ship: your agents in production are yours, including the incident when one misbehaves

What Success Looks Like
First three months: one real workflow runs end to end on the agent runtime in production, with its evals in CI and every step in the audit log.
First six months: more agents are live on the same layer, and at least one of them was built by one of the full stack engineers.
First twelve months: adding AI to a stage of the flow is a few weeks of feature work for any engineer here, and you can show an auditor how any of it reached the answer it did.

How We Work
Linear as the single board: work traced from issue to acceptance criteria to PR, reporting as a by- product of the work
A small team: full stack engineers with a lead, plus AI and platform specialists, on a .NET and Next.js platform on Azure Prototype-led: we design the flow before we build it, against a shared design system
AI in daily use across the team: coding agents in the repo with the context they need, and AI tooling
for research, prototyping and analysis
A customer base we can pick up the phone to, and an implementation team on their sites every week

Tech Stack You'll Use
Backend: C#, .NET 10, Rebus messaging, SQL Server (database per company)
Frontend: Next.js, React, TypeScript
Cloud: Azure (Container Apps, Azure SQL, Service Bus, Front Door), Bicep, GitHub Actions
AI: Model-agnostic by design, with MCP as the tool protocol
Tooling: Claude Code, Linear, GitHub

What We're Looking For     
You've taken multiple LLM-powered systems from prototype to production, at least one of them agentic (multi-step, tool-using, running unattended), and you can walk us through what broke and
what you'd redo Product instincts and a builder's disposition: you can sit with a planner or a quoting manager, translate their workflow into an agent architecture without being handed a spec, then write the code,
ship it, and talk to the people who use it Strong backend and systems engineering experience (5+ years). You've built event-driven systems,
you can explain why the transactional outbox exists, and you think in contracts, idempotency and failure modes
You treat evals and observability as part of the definition of done, and you've wired LLM evaluation and observability tooling
You treat AI security as an engineering problem: injection containment, retrieved content handled strictly as data, and hostile inputs in the eval suite
Clear opinions on build versus buy: you know which parts of an agent stack to take from providers and which parts you have to own
Comfortable being first: setting standards, writing them down, and reviewing others' work against them
Strong written communication: you'll work with a team split across time zones
Preferred Experience
You've shipped something in manufacturing, ERP, supply chain or industrial operations, or you're genuinely curious about how factories run
Document extraction pipelines: structured extraction from PDFs and email, OCR cascades, confidence-scored human review