Principal AI Architect Enterprise AI Delivery

Job Summary: Architecture authority and final escalation. Set the reference architecture for AI systems we deploy into client environments - model selection and routing, retrieval design, tool exposure, orchestration, guardrails, observability, and evaluation. Be the last stop on hard architecture, performance, and deployment calls. Production, not proofs of concept. Own the standards that get prototypes to production: design and code reviews, automated testing, quality gates, release-readiness checks, and integration with clients' real systems of record and identity. Open-weight model engineering and unit economics. We price per unit of output, which makes cost per inference a margin decision. Own what to fine-tune, on what base, with what data, and how to serve it - and the discipline of proving the tuned model beats the frontier baseline it replaces. Then routing, caching, quantization, batching, and distillation on top. Team capability and reusable IP. Lead 30+ engineers and data scientists across concurrent client implementations and build the reference architectures and accelerators that make the second implementation cost materially less than the first. Success is a team that solves hard problems without you. Commercial partnership. Work with sales and consulting on client workshops, proposals, effort and cost estimation, and technical due diligence - assessing feasibility honestly, including when the answer is no. What we are screening for Evidence of production, at scale. You have designed and shipped AI systems that ran in production against real enterprise data and real users, and can walk through the trade-offs that kept them working. Pilots and demos do not count. Fine-tuning open-weight models for production. You have taken an open-weight model, curated the training data, chosen a full fine-tune against LoRA or QLoRA deliberately, evaluated it against the frontier baseline, and served it under real load. Candidates whose entire experience is calling hosted frontier APIs are not a fit for this role. Cost and performance engineering. You have measurably reduced the running cost or latency of an AI system and can describe how, with numbers - quantization, distillation, routing, caching, or serving changes. Evaluation as infrastructure. You treat evals as a first-class engineering concern - offline suites, regression gates in CI, online monitoring, and human review. If observability is where you stop, this role is not a fit. Data readiness and governance. You can look at an enterprise's data estate and say whether it supports the proposed system before the build starts. You have working command of the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act's risk tiers - and, for financial services, model risk management practice as it now stretches to generative and agentic systems. Hands-on depth and team leadership. You still write and review code. Python assumed, plus distributed systems, APIs, containers and Kubernetes, CI/CD, GPU serving, one major cloud, and platforms such as Databricks or Snowflake. You have made senior engineers better and can point to people who outgrew needing you. Baseline: 12+ years across software engineering, architecture, data platforms, and AI/ML, including time in a startup or product company we need engineering velocity, not only enterprise process. Degree in Computer Science, Engineering, or a related technical field. No certifications required or preferred we are interested in what you have shipped.