Skills-Inference Data Platform Architecture Director — Travel Marketplaces

The mandate

The marketplace has three incompatible skill taxonomies, self-declared profiles with low trust and experimental models that infer capability from code, tickets, learning and project histories. Leaders want a common view for reskilling and capacity decisions, while employees cannot see why a skill was attributed or challenge stale evidence. Early pilots reproduce visibility bias and confuse exposure to a task with demonstrated proficiency.

The named deliverable is a Skills-Inference Data Platform Architecture covering legitimate purpose, evidence provenance, ontology mapping, confidence, decay, employee review, correction, consent where required, protected attributes, model evaluation, access, downstream-use controls and human decision boundaries. It must separate discovery suggestions from validated skill claims and prohibit automated employment consequences based solely on inferred profiles.

At week four, milestone one accepts source-purpose and harm maps; week eight approves the canonical evidence model. Week thirteen closes inference and challenge rules, week seventeen completes seven role journeys, and week twenty-two tests three adverse-inference cases. Milestone six at week twenty-six delivers accepted architecture decisions, control specifications, measurement pack, steward playbook and sequenced build backlog.

Acceptance requires employee representatives to trace every displayed skill to permitted evidence, workers to correct an intentionally false inference, and workforce planners to avoid using low-confidence attributes in selection. The Chief Data Officer and Chief People Officer jointly sign when client teams process twelve unfamiliar profiles, a taxonomy change and a source withdrawal without consultant interpretation or silent downstream breakage.

The client will provide taxonomies, source inventories, sample records, model documentation, role structures, decision journeys, access patterns, employee concerns, consultation obligations, privacy assessments, platform constraints and named technical owners. Excluded work comprises production model building, vendor selection, individual profiling, workforce selection, performance rating, legal opinion, collective consultation and operation of the eventual platform.

Why this is external work

Data scientists can optimise predictive accuracy and talent teams can describe intended use, but neither is sufficiently independent to resolve provenance, contestability and employment-boundary questions created by their own pilots. The client needs a finite architecture that works across both disciplines before scaling a consequential workforce data product.

What you will own

• Catalogue every source signal, collection purpose, provenance weakness, population gap and prohibited downstream interpretation.
• Reconcile competing taxonomies through stable concepts, contextual proficiency, evidence strength, expiry and transparent mapping rules.
• Specify inference confidence, abstention, decay, employee visibility, correction workflow and accountable human validation points.
• Design bias tests for opportunity visibility, work allocation, language, tenure, occupational group and missing digital exhaust.
• Connect architecture decisions to privacy, works-council, access, retention, audit and model-change control requirements.
• Rehearse false attribution, withdrawn consent, taxonomy split, proxy discrimination and an unauthorised staffing-system feed.
• Deliver target data architecture, inference policy, harm controls, steward handbook, acceptance evidence and implementation backlog.

Candidate qualifications

• Architected workforce skills, talent intelligence or adjacent human-capital data products across multiple operating countries.
• Distinguished activity traces, self-claims, assessments and manager evidence when expressing contextual proficiency and confidence.
• Built contestable inference with provenance, abstention, expiry, correction and downstream-use constraints visible to employees.
• Evaluated representation and proxy risks where digital exhaust varied by role, language, access or manager assignment.
• Worked credibly across people analytics, data science, privacy, employee relations and worker-representative governance.
• Transferred a platform architecture through client-run false-inference, source-withdrawal and taxonomy-change acceptance scenarios.

Non-negotiables

• Can lead Barcelona laboratories and all three adverse-inference rehearsals inside the six-month delivery window.
• Brings consequential workforce-data architecture; generic recommendation engines or HR dashboards alone are insufficient.
• Will disclose talent-intelligence platforms, data vendors, assessment providers, employers and model-audit relationships.
• Will not profile named employees, select vendors, decide staffing, rate performance, provide legal advice or build production models.

1. 49 words maximum. Describe a workforce skill inference you forced a platform to suppress or qualify.
2. 49 words maximum. How did an employee discover, understand and correct a false inferred capability?
3. 49 words maximum. Which source-withdrawal test would reveal hidden dependence in your proposed architecture?