We are looking for a Senior Machine Learning Engineer to own LLM-powered systems end-to-end - from a first proof of concept built in days, to a production system a client trusts with their core workflows. You will work inside AI Labs, where most engagements start as an ambiguous business problem (complex legal documents, scanned and digital; extraction, classification, assistants, agents) and end as a working system with measurable quality.

This is a hybrid of engineer and scientist. You write production-grade Python and ship it through CI/CD, and you treat model quality as an engineering discipline: you do not say "it works" — you show the eval. Curated datasets, graded criteria, baselines, regression thresholds wired into CI. You know that an LLM feature without evals is an untested feature, and you build the measurement before you build the demo.

You are vendor-agnostic and cost-aware: you choose between a prompt, a RAG pipeline, a fine-tune, a classical model, or plain deterministic code based on the problem, the latency budget, and the cost per request - not on what is fashionable. You use AI tools daily to accelerate your own work, with the judgment to know when to trust the output and when not to.

Because we are a consultancy, you are also client-facing: you demo live, explain trade-offs to non-technical stakeholders, size effort honestly in pre-sales, and turn what you learn into assets the rest of the company can reuse. This is not a role for someone who waits for a well-specified ticket. We need an engineer who challenges the framing of the problem, says when the data does not support the ask, and treats production confidence as a deliverable.

Required Qualifications

• 5+ years of professional software engineering experience, with production ownership — not notebooks handed off to someone else.
• 2+ years building LLM-powered systems in production or in delivered client engagements: prompt engineering, structured extraction, RAG, workflows, or agents.
• Experience designing and running evals: building eval datasets, defining graded criteria, establishing baselines and regression thresholds, and using them as a release gate. Candidates who have only eyeballed outputs will not pass this bar.
• Advanced Python: clean, tested, reviewable code; async, typing, packaging, dependency management.
• Hands-on MCP experience and judgment on when to expose capability as a tool/server rather than hard-wire it.
• Hands-on RAG experience: chunking, embeddings, vector stores, reranking and the ability to explain how retrieval quality was measured.
• Experience with at least one LLM orchestration framework (LangChain/LangGraph, LlamaIndex, Pydantic AI, or equivalent) and the judgment to work without one.
• Cloud deployment of AI/ML workloads, preferably AWS (Bedrock, SageMaker, Lambda, ECS, S3) or demonstrable equivalent on GCP/Azure.
• CI/CD and engineering hygiene: Git, automated tests, reproducible builds, containerization; ability to ship without a platform team holding your hand.
• Practical understanding of cost and latency engineering for LLM systems: token accounting, caching, model routing, streaming.
• Awareness of LLM security and privacy risks: prompt injection, data leakage, PII handling in a client-data context.
• AI-first working style: daily hands-on use of AI coding tools (Claude Code, Cursor, Copilot, or similar) with the judgment to know when to trust the output.
• Senior judgment: operate autonomously, challenge an ill-posed problem, prioritize by risk and business impact, own outcomes.
• Excellent written and spoken English; comfortable in a fully remote, distributed, client-facing team.

Preferrable:
- Experience with an eval or LLM-observability framework (LangSmith, Langfuse, Weave, Braintrust, Arize Phoenix, DeepEval, Ragas, or similar).
- Experience building LLM-as-judge graders and validating judge/human agreement.
- Fine-tuning / adaptation experience (LoRA, SFT, distillation) and a clear view of when it pays off versus prompting or RAG.
- Hands-on deep learning: PyTorch, with models trained on real projects.
- Classical ML and NLP depth: classification, NER, embeddings, recommendation, computer vision, time series.
- MLOps tooling: MLflow, Weights & Biases, DVC, model/prompt versioning and registries.
- Serving open-weight models (vLLM, Ollama, Bedrock/SageMaker endpoints) and reasoning about GPU cost.
- Voice / real-time interfaces (STT, TTS, streaming agents).
- Legal, financial, or other regulated-domain document experience.
- Consulting or client-facing delivery experience; live demos, discovery, pre-sales.
- Public evidence of craft: open-source contributions, papers, talks, or a track record of writing about applied AI.