Job Description
Experience: Senior-level capability, demonstrated through work rather than years of experience. Candidates at any career stage are welcome.
Help build the tools and data architecture that make client's AI/ML systems effective, efficient and trustworthy.
This role has two entry points:
AI development tooling - our primary focus: Build checks that catch conceptual errors and architectural drift in AI-assisted code.
Data architecture: Establish what clinical data exists, what can be reliably linked and legitimately used, and how it should be modeled.
You do not need both. We would rather hire someone excellent in one area and curious about the other. We'll shape the engagement around your strengths.
What matters most
Three capabilities matter more than familiarity with our stack:
Learning unfamiliar domains. You can enter a subject you don't know, identify the important questions, and become useful. You recognize and acknowledge when your initial understanding was wrong.
Reasoning from evidence. You distinguish measurements from inferences, proposals from decisions, and structural possibility from demonstrated capability. You state what the evidence does-and does not-establish.
Making abstractions useful. You turn grammars, contracts, invariants, type systems, or measurement definitions into tools and models others can use.
Track A: AI development tooling
AI assistants accelerate code generation, but they fail in recurring ways. We plan for those failures rather than hoping to avoid them: the checks that enforce requirements and architectural constraints live outside the model, so they hold regardless of which assistant produced a change.
You'll work on:
Architecture checks: Compare source-tree versions and report changes across distinct dimensions, with supporting evidence rather than a single score.
Static analysis: Analyze Python code for duplication, dependency cycles, boundary violations, exception handling, and other structural risks.
Reproducible, fail-closed tooling: Produce identical reports from identical inputs. A check that cannot run must never count as a pass.
Decisions wired to checks: Connect architecture decision records to the gates that enforce them, with verifiable identities, statuses, and references.
Schema-driven generation: Generate types, constants, and validators from a single source of truth, with golden fixtures that record expected behavior.
Evaluation beyond mechanical checks: Build reproducible model-graded probes and audits that make their limitations-and unevaluated cases-explicit.
Track B: Data architecture
Our clinical data platforms combine many upstream systems. Even when sources use the same schema, they can differ in which fields contain data, whether records can be linked reliably, and how the data may be used.
You'll work on:
Source characterization: Separately establish what exists, what is populated, what can be linked, and what may be used.
Table contracts: Create versioned, per-table contracts that generate reference views, inventories, coverage reports, and explorers.
Reproducible measurement: Define grain, time windows, deduplication, exclusions, and unresolved cases; check that populations and denominators are comparable before comparing results.
Safe analysis at scale: Estimate query costs before accessing multi-terabyte tables and return aggregates or metadata rather than records.
Modeling and reuse: Assess existing components against requirements, reason carefully about cross-system identity, and evaluate logical models, storage, and serving as related but separate decisions.
How we work
We use AI agents throughout development. You will work across roles - product owner, architect, and engineer - while keeping requirements, design decisions, and implementation responsibilities distinct. People retain responsibility for approval and authorization.
Define before implementing. Work proceeds from product requirements and product behavior through high-level design, detailed specifications, and an implementation plan. Documentation is proportional to the change.
Make decisions explicitly. Record product requirements and behavior in the product documents, and technical design decisions in architecture decision records. Detailed specifications and implementation plans describe those decisions; they do not introduce new ones. Agents may investigate and recommend, but cannot approve or authorize.
Match review to purpose. People review requirements, product behavior, and high-level design directly. Detailed specifications and implementation plans are written primarily for agents and authorized by people based on review and verification evidence.
Check facts as well as instructions. Independent reviews, source checks, and executable tests examine whether instructions are clear, claims are supported, and results meet requirements. Agreement between documents is not enough.
Resolve questions before proceeding. Missing decisions, conflicting evidence, and unclear instructions return to the responsible owner. Affected work stops until the instructions are corrected or a temporary deviation is explicitly authorized.
Verify completion. Implementation plans state the required results and checks. Completed work is reviewed independently, and completion requires recorded evidence that all required checks passed.
What you bring
Strong production Python skills. For Track A, experience with Python's standard library and ast or equivalent code-as-data work.
For Track B, strong SQL and analytical reasoning, especially around grain, joins, and denominators.
Clear technical writing. Design documents, assessments, and decision records are core deliverables.
Independent delivery and sound judgment, including recognizing when a question requires an explicit design decision or approval from someone else.
Evidence of substantial self-directed learning followed by practical application.
Fluency with AI coding assistants, paired with the judgment to challenge plausible but incorrect or architecturally damaging output.
No degree, minimum number of years of experience, veterinary or ML background, or experience with our exact stack is required.
Helpful, not required
Relevant experience might include:
Static analysis, linters, codemods, architecture-rule tooling, or schema-driven code generation.
Property-based, golden, mutation, or contract testing.
Model-graded evaluation, probe design, or measurement of nondeterministic systems.
Data modeling, catalogs, lineage, record linkage, dbt, Spark, or schema languages.
Declarative rules, DSLs, constraint systems, program analysis, or information theory.
Depth in one area plus strong reasoning is enough to apply.
What success looks like
First 30 days: Ship a small, useful check, measurement, or contract, with documentation that makes it reviewable.
By 90 days: Own a defined area and become a trusted resource for it.
By the end of the engagement: Leave work that is adopted, documented, and maintainable by the team.