Senior AI Data Scientist — Agentic Process Development

team.blue Global

About The Position

We are looking for a Senior AI Data Scientist to streamline HR processes at team.blue — not by analysing them, but by building agentic systems to run them. Recruitment, onboarding, performance, rewards and offboarding are each multi-step processes spanning several systems and up to 25 countries, and your mandate would be to create systems that can streamline them end to end. The method matters more than the domain: map a process, quantify what it costs in headcount, score which steps an agent could take, build a proof of concept, and take it to production. This work sits closer to building autonomous, side-effecting systems than to building predictive models. The agents you design would be able to revoke IT access, issue signed contracts, and flag pay outliers into approval workflows. A wrong output here is not a bad number someone can catch — it is a high impact action taken in the world.

Requirements

7+ years building data and ML systems in industry, spanning both sides of the LLM shift.
Shipped something that had permission to take an irreversible action affecting real customers.
Expert in Python and ML.
Ship end to end: Python someone else can still read in six months, a current toolchain (uv, Docker or an equivalent), your own container, your own instrumentation.
Production experience with multi-step, tool-calling LLM workflows — orchestration, retries, idempotency, timeouts, partial-failure recovery.
State-machine design, not only train/serve pipelines.
Cost and latency engineering as a first-class concern — model routing, caching, batching, and the instinct to know what a flow costs per run before Finance asks.
A safety instinct for systems that take actions — staging modes, approval gates, least-privilege scoping, rollback.
Evaluation design for generative and agentic output — LLM-as-judge, golden-transcript regression suites, red-teaming.
Applied statistics you can adjudicate with.
Process mapping and quantification — you can sit with a process owner, capture what actually happens rather than what the policy says, and attach a number to it.
Facilitation — you can run a workshop with senior non-technical stakeholders and leave with requirements.
Executive-grade written business cases — cost modelling and framing for a Finance audience, which is a different skill from data storytelling.
Technical vendor evaluation — judging an HR-tech vendor on API surface, data model, extensibility and true integration cost, not on the sales deck.

Nice To Haves

Master's or PhD in Computer Science, AI, Machine Learning or a related field
Existing EU AI Act / high-risk-AI-system familiarity
PromptOps at scale — versioning, testing and rollback of prompts as production artefacts
DataOps/MLOps practice: deploying and monitoring models and pipelines
Prior exposure to HRIS/HR-tech, or to automation in any audited or regulated domain

Responsibilities

Map a process, quantify what it costs in headcount, score which steps an agent could take, build a proof of concept, and take it to production.
Create systems that can streamline HR processes end to end.
Design state transitions: what triggers, what branches, which systems get called, where it waits, when it escalates, and what happens when step 4 of 9 fails.
Build guardrails before the capability — dry-run mode, an approval gate ahead of anything irreversible, least-privilege scoped credentials, a rollback path.
Decide where a human stays in the loop, at what confidence threshold, and design a review queue they will actually use.
Write evals for output that precision and recall do not capture: task-completion rate, hallucination rate, gendered or culturally biased language in AI-drafted reviews.
Wire an agent to a webhook instead of a nightly batch pull — and making the handler idempotent so a retry does not offboard someone twice.
Decide which steps in a flow warrant a frontier model and which can run on something cheap, then proving that routing decision with numbers.
Sit in a vendor demo asking what their API actually exposes, what their data model looks like, and what integration really costs us.