← Back to Work

EU AI Act · Bias Testing · Case Study

Rediscovering talent already in the pool

A crawler-and-alert system for surfacing overlooked candidates already in the talent pool — designed to meet EU AI Act high-risk obligations from day one, not retrofitted later.

ROLE

Sole Designer — with CTO & CEO

TIMELINE

Jun – Oct 2026

STATUS

In progress — paused by company, continuing independently

Context

The Problem

Recruiters are busy, and often skip checking their existing talent pool when filling a role — so they miss talents already on file who would be a great fit.

Wait a sec — a different problem

Partway through scoping this as a recruiter-workflow feature, a second question surfaced: is this actually a different problem? The EU AI Act applies to any system that assists employment decisions — if built, it needs to conform from day one, not get retrofitted after. That catch is what shifted this from a feature build into a high-risk AI design problem before any design work began.

What this requires

Because this touches Annex III employment decisions, the system needs to already conform to EU AI Act high-risk obligations — transparency, human oversight, bias testing — not have them bolted on after the fact.

Goals

Make talents already in the pool visible

Decrease time recruiters spend searching

Increase outgoing job links

Increase hires sourced from the pool

Process

Solutions considered

AI-based matching

Match customer job postings against pooled talent profiles directly.

Tradeoff: broader accuracy potential, but a bigger compliance surface — more automated decisioning to test and document.

Low-tech: email campaigns

Send talents postings matching their past searches. Limited — only works if a search was made and saved.

Tradeoff: simplest to build, but only reaches talents who already searched and saved a role.

Crawl & alert

Crawl the customer's job postings; alert on matches via email and in-app (Talent Explorer).

Tradeoff: keeps a human in the loop by design (alerts, not auto-decisions), but needs ongoing crawler upkeep.

What we built

Recruiters set keyword filters (e.g. "Sales") to scope which postings to track. → input specification

The system crawls the customer's career site and analyses postings for required skills, experience, salary, languages, and education. → explainability

Matching talents from the pool are sent via a configurable-frequency email digest, and surfaced in the global activity stream. → accuracy disclosure

Bidirectional filtering: recruiters can filter talents by keyword, or select a job and see which pooled talents fit it. → human oversight

Bridging HAX guidelines to the EU AI Act

A concrete link between CS3 (my MAS thesis on Microsoft's HAX guidelines) and this case study's obligations — the first time these two projects connect directly.

Input specification

Demonstrates possible inputs up front. (HAX G1 PD, G11 PD/PF/PG)

Explainability

Shows outputs and maps inputs/behaviors to results, with example-based explanations. (HAX G1 PE, G11)

Accuracy disclosure

Makes clear what the system can do, and how well it does it. (HAX G1, G2)

Human oversight

Supports efficient dismissal, correction, undo, and batch-editing. (HAX G8, G9)

Testing for bias

Bias is a real concern when an AI model selects matching candidates for a position. Testing needs to include talents across demographics, including minorities — comparing how the model matches candidates on anonymised data (age, gender, name removed) versus full data, to reveal bias in the model itself. Processing age and gender for this test is itself special-category data under GDPR Article 9 — a second compliance surface beyond the AI Act, needing its own lawful basis before any test data is touched, not just an AI Act checkbox.

Results

Impact

Created clarity around the competitor landscape, feasibility, and legal/regulatory obligations — the groundwork needed before this could move from concept to build. The four goals above remain unmeasured — the initiative was paused before reaching a testable state; they're the criteria this will be evaluated against if it resumes.

"Wait a sec — is this actually a different problem?"

The moment that reframed the whole project — from a recruiter feature into a high-risk AI system, before any design work began.

Learnings

Bridging HAX guidelines to AI Act obligations gives a repeatable framework for future high-risk AI design work — not a one-off for this case study alone.