AI/ML Engineer

Plugging into an LLM API is not the same as building an AI system that works in production under real-world pressure. The gap between those two keeps widening, and the engineers who can bridge it are becoming the most valuable people on any team. These are the people who understand model behaviour, evaluate rigorously, and design systems where AI is a load-bearing component.

We're not looking for engineers who have just discovered GenAI. We're looking for engineers whose craft is AI: who reason about model behaviour the way strong software engineers reason about distributed systems, who build evaluations before they build features, and who can tell you exactly why a model is failing and how to fix it.

This role is for the engineers our clients rely on when the stakes are highest: when the AI has to actually work, not just demo well.

Concretely, you will:

• Design and build production AI systems where model behaviour is load-bearing: retrieval architectures, agentic workflows, multi-model orchestration, structured generation pipelines.
• Own the evaluation methodology end-to-end. Define what "good" means for a given system, build regression suites, run offline and online experiments, and track drift and degradation once it's in production.
• Make model and architecture decisions with real trade-offs in mind (latency, cost, accuracy, data sensitivity, operational complexity). Know when to use RAG, when to fine-tune, when to route across models, and when AI isn't the right answer at all.
• Debug AI systems at the model level: prompt behaviour, retrieval failure modes, tokenisation edge cases, and generation drift. Build the observability needed to catch these before they reach users.
• Contribute to or drive LLMOps infrastructure: inference serving, caching, monitoring, experimentation, and cost control.
• Translate ambiguous business problems into concrete AI system designs, with clear reasoning about what's feasible today versus what's still speculative.
• Stay at the frontier of AI/ML research: read papers, evaluate techniques critically, and bring what's genuinely useful into practice.
• Use AI coding tools (Claude Code, Cursor, Copilot, or similar) as a natural part of your workflow. You, more than anyone, understand what these tools can and can't do.

Key Requirements

• 3+ years as a software engineer and/or at least 2+ years of substantial production work on AI/ML or GenAI systems.
• Deep understanding of modern AI systems: LLMs, embeddings, vector search, retrieval pipelines, agentic architectures, evaluation frameworks, and production observability for AI.
• Working knowledge of classical ML fundamentals: supervised learning, evaluation metrics, overfitting, data distribution shifts. You know when a transformer isn't the right tool.
• Strong engineering fundamentals in Python (primary), ideally with working comfort in TypeScript/Node.js.
• Proven ability to design and own evaluation methodology, not just use existing frameworks. You can tell signal from noise in a messy eval.
• Can reason about model behaviour at depth: why a system failed, how to measure improvement, how to build systems that degrade gracefully.
• Comfortable at the boundary between research and production. You read papers, assess techniques critically, and decide what's worth adopting.
• High agency and technical leadership. You make calls, communicate trade-offs clearly, and bring others along.
• Pragmatic about when AI is the right tool, and when it isn't. You'd rather ship a simpler non-AI solution when that's genuinely the better answer.
• Low ego, high standards. You hold the bar on technical rigour without slowing everyone down.

Nice-to-Have

• Hands-on experience with fine-tuning, LoRA/QLoRA, DPO, or similar adaptation techniques, with a clear sense of when fine-tuning is worth the cost.
• Has built evaluation harnesses from scratch: LLM-as-judge pipelines, regression suites, red-teaming setups, or human-in-the-loop eval workflows.
• Experience deploying AI systems under serious latency or cost constraints.
• Contributions to or deep familiarity with the open-source ML ecosystem (HuggingFace, vLLM, LangGraph internals, MCP, agent protocols).
• Experience with the full LLMOps lifecycle: model routing, inference serving, caching, monitoring for drift, cost attribution.
• Can discuss specific papers (Chinchilla, ReAct, DPO, Toolformer, etc.) with informed opinions, not just summaries.
• Has worked on classical ML problems (recommendation, forecasting, NLP pre-LLM) and can speak to what changes with GenAI and what doesn't.
• Active contributor to the AI tooling or research community (papers, open-source, blog posts, talks).
• Has shipped real products using AI-assisted coding workflows and can speak to how it changes the rhythm of deep AI/ML work.

PX interview → Technical interview → Final conversation → Offer
* The technical interview focuses on real problems, not algorithmic puzzles. 

We evaluate candidates holistically. If you don't meet every requirement listed above, apply anyway. We care more about how you think and work than whether your CV matches a checklist.