Principal AI Engineer

About this role

Principal AI Engineer — Velsera
AI Platform & Enablement · Reports to the CTO · Senior individual contributor
About the role
Velsera builds software and infrastructure for precision medicine — research platforms, clinical and diagnostic applications, and the systems that keep them running in regulated environments. We are adding AI capability across that portfolio and inside our own operations: governed model access, self-hosted and managed LLM serving, evaluation and audit, and integration into the products and business processes people already depend on.
This is a deliberately broad role. You will be deployed where the highest-value AI work is at the time — a customer-facing product capability in one quarter, an internal enterprise workflow in the next, a strategic account or funded program after that. The mandate stays the same wherever you land: design and ship production AI systems that hold up under real compliance requirements, work across AWS, Azure, and GCP, and leave behind reusable patterns rather than one-off builds.
You will set technical direction for what is expected to grow into an AI platform and enablement team.
What you'll work on

• Build a governed model access layer — self-hosted open-weight models, cloud-managed models (Bedrock, Vertex AI, Azure OpenAI), and customer- or partner-supplied models — designed to be consumed by more than one product or business function.

• Integrate AI capabilities into product experiences and enterprise workflows across batch, interactive, and agentic patterns.

• Establish the patterns everyone else reuses: evaluation, versioning, approvals, audit trails, cost control, guardrails, and safe rollout and rollback.

• Partner with product, engineering, security, QARA/compliance, IT, and scientific and commercial teams to introduce AI-native architectures that people can actually adopt.

• Move between assignments as business priorities shift, and make what you build in one part of the business usable in the next.
What you'll deliver (first 6–12 months)

• A production-ready, compliant AI/LLM serving and invocation layer that at least two products or business functions adopt — multi-tenant, auditable, and secure.

• A model governance workflow (intake, evaluation, approval, versioning, deprecation) that satisfies both regulated customers and our own quality system.

• Two or three AI capabilities shipped end to end in different parts of the business — for example a customer-facing product feature, an internal process automation, and assisted validation or compliance tooling.

• Integration patterns that preserve reproducibility, traceability, and standards alignment wherever the work lands.

• Operational readiness: monitoring, evaluation harnesses, incident playbooks, cost visibility, and measurable SLOs for key AI services.

• A defensible internal point of view on where we should build, buy, or not use AI at all — backed by what you shipped.
How we build (and what we'll expect you to optimize for)
You will make trade-offs in an environment that is multi-product, multi-cloud, standards-driven, and compliance-heavy. A few things matter a lot here:

• Reusability over one-offs. Design the second and third use case into the first one. A solution that only works for one product or one team is a partial solution.

• Standards and clean interfaces. Prefer open standards and well-defined boundaries over bespoke integrations.

• Multi-cloud, multi-deployment reality. AWS, Azure, and GCP are all in play, alongside customer-managed and self-hosted environments. Avoid hard dependencies on a single provider's AI stack.

• Security and auditability by default. Access control, logging, traceability, and data governance are part of the design, not add-ons.

• Reproducibility. AI features have to fit into workflows and processes that need to be repeatable and explainable, sometimes years later.

• Proportion. Ship the smallest thing that genuinely works, then harden it. Governance that makes a workflow unusable has failed.
Requirements
What we're looking for
Must-haves

• 7+ years in software engineering, including 3+ years shipping AI/ML systems to production.

• Strong Python, plus one of Java, Go, or TypeScript; comfortable in a polyglot codebase and in production code review.

• Hands-on experience with secure cloud architecture on at least one major cloud — network isolation, IAM boundaries, private connectivity, audit logging — and readiness to work in the others.

• Experience operating or integrating model serving across delivery modes: self-hosted open-weight models, managed model APIs, and customer-provided models.

• MLOps/LLMOps experience with tooling such as AWS Bedrock, Google Vertex AI, Azure AI Foundry, or equivalent.

• Built governance for ML/LLM systems: evaluation, versioning, approvals, rollout and rollback, deprecation.

• Comfortable designing for regulated or audited environments (HIPAA, 21 CFR Part 11, GxP, FedRAMP, SOC 2, GDPR, or similar).

• Experience with RAG and LLM tool-use/agentic patterns beyond prototypes, including how you evaluated them.

• Track record integrating with systems you don't own — existing products, third-party SaaS, enterprise data sources — without breaking them.

• Clear written communication for mixed audiences: engineering, product, security and compliance, business stakeholders, and scientists.

• Comfort switching context across problem domains and starting from ambiguous requirements.
Nice-to-haves

• Experience in genomics, biomedical data, or life sciences platforms.

• Software built under a quality management system (ISO 13485, IEC 62304, IVDR) or in clinical/diagnostic contexts.

• Integrating AI capabilities into workflow or orchestration engines (CWL/WDL/Nextflow or similar).

• Familiarity with GA4GH standards (WES/DRS/TRS) and/or clinical data models (FHIR, OMOP).

• Enterprise systems integration: CRM, ERP, ITSM, document and quality management, collaboration suites.

• Production…