Non-negotiable. This is what we screen on first, and there is no AI in it.

• Data modelling across SQL and NoSQL, and a clear view of where each one breaks

• Event-driven architecture in production with Kafka, RabbitMQ or equivalent. Ordering, replay, idempotency, dead-letter handling

• Distributed systems failure modes: partial failure, retries, backpressure, timeouts, circuit breakers

• Caching strategy and invalidation, at a scale where getting it wrong hurt

• Observability. You have instrumented a system, not just read someone else's dashboard

• At least one cloud deeply. Not three superficially

• API and integration patterns against real enterprise surfaces such as ERPs, CRMs and data platforms

• CI/CD, containers, release engineering

• You have owned something in production. On-call, incidents, the whole thing

Agent systems are distributed systems with a non-deterministic component in the middle. If the fundamentals are not there, the AI layer does not survive contact with production.

Two capabilities that do not come free with an engineering background. These are the ones we probe hardest.

Evaluation and Error Analysis

• You have built an eval set for a non-deterministic system and used it to make a decision

• You can decompose a failure: is this a model problem, a retrieval problem, a data problem or a product problem?

• You debug agent behaviour systematically, with traces and evidence, not by gut feel

Context Engineering

• Deliberate management of what enters the model on each turn, and why

• Memory strategy: what persists across steps, across sessions, and what should not

• Tool surface design: what the agent can reach, how it is scoped, what happens when a tool fails

• Cost and latency as first-class design constraints, not something discovered in the invoice

Equal weight to the engineering. This is a client-facing role in a services business.

• Carry a market point of view on models, tooling and delivery patterns, and translate it into what it means for a specific client, in their language

• Run discovery workshops, solution reviews and delivery cadences with client teams

• Advise clients on standing up their own AI capability: practice structure, operating model, skills, governance, build versus buy, and adoption across their engineering org

• Shape and defend technical proposals, PoC plans and roadmaps. Own the story end to end

• Translate business problems into architectures for CXO-level stakeholders without hiding behind jargon

• Say no when the answer is no. Tell a client when agents are the wrong tool, and be able to explain why

The work itself.

• Own end-to-end delivery of AI-native programs, from architecture through production

• Design multi-agent orchestration using LangGraph, CrewAI or equivalent. Agent topology, tool routing, memory, state, fallback and recovery paths

• Integrate agent systems with enterprise systems of record, not toy datasets

• Build RAG where RAG is the right answer, from chunking through retrieval, re-ranking and evaluation

• Run agentic coding workflows with Claude Code, Cursor, Codex or equivalent, and lead projects where AI writes significant portions of the codebase while you guide, review and ship it

• Work with MCP and shared-context tooling. Design what agents are permitted to reach and under what controls

• Bring governance into the design from day one: agent identity, privilege scoping, audit, human-in-the-loop, kill switches

• Contribute to reusable frameworks and accelerators the wider practice can use

• Build AI engineering capability across the delivery organisation through mentoring, standards and review quality

• Evaluate new models, frameworks and tooling before the hype catches up, and kill the ones that do not earn their place

• Contribute to internal knowledge bases and practice assets

• Someone who lists LLMs on a resume but has only called the API in a notebook

• People who explain everything in terms of frameworks they have never deployed

• Consultants who can only narrate what others have built

• Engineers who cannot hold a client conversation, and client-facing people who cannot build

Stage 0, System design with no AI. A distributed systems problem with no LLM in it. Event flow, data model, failure handling, scale. Twenty-five minutes.

Stage 1, Technical screen. Walk us through a system you built. Bring an artifact: an eval set, a trace from a run that failed, a postmortem, a cost-per-task number. Something real, not slides.

Stage 2, Architecture discussion. A business problem, live. Design the solution and defend the trade-offs, including the decision not to use agents.

Stage 3, Client simulation. Present your approach to a non-technical executive audience, then advise them on how to build the capability internally.

Non-negotiable. This is what we screen on first, and there is no AI in it.

• Data modelling across SQL and NoSQL, and a clear view of where each one breaks

• Event-driven architecture in production with Kafka, RabbitMQ or equivalent. Ordering, replay, idempotency, dead-letter handling

• Distributed systems failure modes: partial failure, retries, backpressure, timeouts, circuit breakers

• Caching strategy and invalidation, at a scale where getting it wrong hurt

• Observability. You have instrumented a system, not just read someone else's dashboard

• At least one cloud deeply. Not three superficially

• API and integration patterns against real enterprise surfaces such as ERPs, CRMs and data platforms

• CI/CD, containers, release engineering

• You have owned something in production. On-call, incidents, the whole thing

Agent systems are distributed systems with a non-deterministic component in the middle. If the fundamentals are not there, the AI layer does not survive contact with production.

Two capabilities that do not come free with an engineering background. These are the ones we probe hardest.

Evaluation and Error Analysis

• You have built an eval set for a non-deterministic system and used it to make a decision

• You can decompose a failure: is this a model problem, a retrieval problem, a data problem or a product problem?

• You debug agent behaviour systematically, with traces and evidence, not by gut feel

Context Engineering

• Deliberate management of what enters the model on each turn, and why

• Memory strategy: what persists across steps, across sessions, and what should not

• Tool surface design: what the agent can reach, how it is scoped, what happens when a tool fails

• Cost and latency as first-class design constraints, not something discovered in the invoice

Equal weight to the engineering. This is a client-facing role in a services business.

• Carry a market point of view on models, tooling and delivery patterns, and translate it into what it means for a specific client, in their language

• Run discovery workshops, solution reviews and delivery cadences with client teams

• Advise clients on standing up their own AI capability: practice structure, operating model, skills, governance, build versus buy, and adoption across their engineering org

• Shape and defend technical proposals, PoC plans and roadmaps. Own the story end to end

• Translate business problems into architectures for CXO-level stakeholders without hiding behind jargon

• Say no when the answer is no. Tell a client when agents are the wrong tool, and be able to explain why

The work itself.

• Own end-to-end delivery of AI-native programs, from architecture through production

• Design multi-agent orchestration using LangGraph, CrewAI or equivalent. Agent topology, tool routing, memory, state, fallback and recovery paths

• Integrate agent systems with enterprise systems of record, not toy datasets

• Build RAG where RAG is the right answer, from chunking through retrieval, re-ranking and evaluation

• Run agentic coding workflows with Claude Code, Cursor, Codex or equivalent, and lead projects where AI writes significant portions of the codebase while you guide, review and ship it

• Work with MCP and shared-context tooling. Design what agents are permitted to reach and under what controls

• Bring governance into the design from day one: agent identity, privilege scoping, audit, human-in-the-loop, kill switches

• Contribute to reusable frameworks and accelerators the wider practice can use

• Build AI engineering capability across the delivery organization through mentoring, standards and review quality

• Evaluate new models, frameworks and tooling before the hype catches up, and kill the ones that do not earn their place

• Contribute to internal knowledge bases and practice assets