Applied AI Architect

Required

15+ years spanning software engineering and quantitative analysis. This role needs both; a strong background in one and a passing
acquaintance with the other will not carry it.
• Production experience with LLM applications: prompting, tool and function calling, context management, evaluation, and knowing
where models fail in practice.
• Experimental design and causal inference — randomized and quasi-experimental designs, difference-in-differences, instrumental
variables, hierarchical models — and the judgment to say when a design does not support the claim being asked of it.
• • Strong Python and SQL, with a statistical stack (pandas, statsmodels, scikit-learn, or R). Data collection: instrumenting and extracting from operational systems and APIs, designing sampling that survives scrutiny, and
knowing when a source cannot answer the question being asked of it.
• Aggregation: resolving identity across systems, joining sources never designed to be joined, and modeling the summary tables
reporting reads from. You need not own the pipeline, but you must be able to build one when the answer depends on it.
• Analytics: exploratory analysis, distributions rather than averages, cohort and time-series work, and reports that state their own
coverage and limits.
• Real familiarity with the software delivery lifecycle — code review, CI/CD, test strategy, release and change management —
sufficient to hold a credible conversation with the teams you are measuring.
• Git and GitLab at instrumentation depth: merge request and pipeline data models, diffs and SHAs, what merge, squash, rebase,
and cherry-pick do to line-level analysis, and the API and hook surfaces available for capturing it.
• Jira and Confluence integration experience — REST APIs, changelog and page version history, the GitLab–Jira development panel,
and the field and label conventions that determine whether the resulting data means anything.
• Care with personnel-adjacent data: aggregate reporting by default, and a clear sense of what should not be built even when it is
technically easy.
• Communication that works in both directions — an executive audience that wants a number, and engineers who will dispute it