Principal AI Architect -Enterprise AI , Newark, NJ (Hybrid)

Location: Newark , NJ (Hybrid-3 days in a week)

Mandatory - Team management : 5 to15+ AI engineers and data scientists Skill required - Open source AI, Open-weight model, LoRA or QLoRA, Python, NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act’s risk tiers , Databricks or Snowflake, Kubernetes , generative and agentic production work , Why this role exists Most enterprise AI never leaves the pilot. Roughly 80% of initiatives fail to deliver their promised value, and only about one in seven organizations has scaled an agentic system to real operational use. The failure is almost never the model. It is curated pilot data that does not survive production, an integration layer underestimated by an order of magnitude, and evaluation infrastructure that was never built. We close that gap inside our clients’ environments at a cost per unit of output that stays defensible as volume grows. That last part is why this is not a frontier-API orchestration job: the economics only work when someone can fine-tune and serve open-weight models where they beat a hosted call. This is the most consequential technical hire on our team. What you will own 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 assumes, 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 processes. Degree in Computer Science, Engineering, or a related technical field. No certifications requiredpreferred — we are interested in what you have shipped. What this role is NOTWe have received strong candidates for adjacent roles who are not right for this one. To be explicit: Not this role WhyFrontier-API integrator - If every system you built was a prompt and a hosted API call, you have not done the engineering this role requires. Data architect - Warehouse, lake house, and pipeline design is a prerequisite here, not the job Cloud / platform architect - Cloud-native infrastructure depth without model-level engineering is not a fitML engineer, classical only- Predictive-modelling depth without generative and agentic production work is insufficient Delivery or engineering manager - If you have stopped writing and reviewing code, this is the wrong role regardless of seniority