AI Recruitment Screening Platform
The client was a fast-growing HR technology vendor serving mid-market companies running roughly two thousand active requisitions across their customer base at any time. Recruiter productivity was the constraint on customer growth — each recruiter handled a maximum of about ten requisitions at the volume of applicants the platform generated, and resume screening took the largest single block of recruiter time. Customers were beginning to ask whether the platform could shortlist candidates rather than just collect them.
Productivity & SaaS Tools
HRTech / Recruitment
13 weeks from kickoff to general availability across customer accounts
5 specialists
The full story
The practical problem was that resume screening at the volume customers needed required either dramatically more recruiter time or AI assistance, and customers were also extremely sensitive to bias concerns. Naive resume-scoring systems had been publicly criticized for amplifying historical bias, and the client’s customers needed both the productivity gain and a defensible position on fairness. The combination of speed and fairness was the design problem, not either one alone.
We built a candidate screening platform that scored and ranked applicants against the job description using semantic similarity over role-relevant signals (skills, experience type, outcomes), explicitly excluded protected and demographic-correlated features from the scoring path, and ran a bias-check module that audited the resulting shortlist for demographic disparities before recruiter handoff. Every score carried a per-candidate explanation, and the bias-check produced a summary that customers could include in fairness disclosures.
What shipped was a recruiter workspace where applicants arrived against a requisition, scored and ranked within seconds, with per-candidate rationale and an automatically-generated fairness summary across the shortlist. Recruiters spent their time on the top of the funnel and on interviews rather than on initial screening, customer time-to-hire dropped substantially, and the bias-check disclosures gave customers a defensible position they could share with their own compliance teams.
Recruiters could not screen at the volume customers generated, and naive AI screening risked amplifying bias in a public way.
Recruiter productivity was the constraint on customer growth, with each recruiter capped at roughly ten active requisitions.
Resume screening took the largest single block of recruiter time per requisition, with most candidates rejected within seconds of review.
Customer demand for AI-assisted shortlisting was rising, but customers were also sensitive to fairness concerns from prior industry incidents.
Naive scoring systems risked amplifying historical bias by learning from biased training data, creating reputational and legal exposure.
No defensible fairness disclosure existed in any competing tool, so customers could not represent the screening process to their own compliance teams.
How we structured the engagement
Designed scoring to be productive and fairness-checkable from day one, not productive first with fairness bolted on.
- 01Phase 01Weeks 1-3
Discovery
Reviewed two thousand recent resumes against a sample of requisitions to taxonomize the signals recruiters actually used. Worked with the client’s legal and DEI leads on fairness requirements and disclosure obligations. Output: a role-relevant signal schema, explicit feature exclusions for protected and correlated attributes, and a bias-check specification.
- 02Phase 02Weeks 4-5
Architecture
Designed a semantic-similarity scorer over skills, experience type, and outcomes, with explicit exclusion of name, address, school prestige, and other demographic-correlated features. Built a bias-check module that audited shortlists for demographic-disparity flags before recruiter handoff. Used FastAPI for the scoring service and PostgreSQL for audit history.
- 03Phase 03Weeks 6-11
Build
Shipped the scorer and the per-candidate rationale generator first, then the bias-check module. Implemented the recruiter workspace with shortlist view, rationale per candidate, and the fairness summary across the shortlist. Built the audit history layer for customer compliance review with per-decision provenance.
- 04Phase 04Weeks 12-13
Launch
Rolled out to ten customer accounts across four industries for four weeks, with recruiter approval required for every shortlist before candidate outreach. Monitored bias-check outputs and recruiter overrides daily, refined the scorer against override patterns. Promoted to general availability after fairness summary review by customers’ legal teams cleared.
What we built, component by component
- 01
Job description parser
Extracts role-relevant signals from the requisition — skills, experience type, outcomes — for use as scoring criteria.
- 02
Candidate semantic embedder
Embeds applicant resumes against role-relevant signals only, excluding protected and demographic-correlated attributes.
- 03
Scorer and ranker
Combines semantic similarity with role-fit heuristics into a per-candidate score with rationale linked to specific resume content.
- 04
Bias-check module
Audits shortlists for demographic-disparity flags before recruiter handoff, generates a disclosure summary per shortlist.
- 05
Recruiter workspace
Shortlist view with per-candidate rationale, side-by-side comparison, and the fairness summary across the shortlist.
- 06
Audit history
Per-decision provenance log including signals used, score, and override actions for customer compliance review.
A requisition is parsed into role-relevant signals and applicants are embedded against those signals with protected attributes excluded. The scorer ranks candidates with rationale per score, the bias-check module audits the shortlist for disparity flags before handoff, and the recruiter reviews on a single workspace with the fairness summary available. Every decision logs to the audit history for compliance review.
The trade-offs we made and why
Excluded protected and correlated features from scoring
Including demographically correlated features even indirectly would have undermined the fairness story. Excluding name, address, school prestige, and similar attributes by design produced a scoring path that customers’ compliance teams could review and approve, which was the precondition for adoption.
Built bias-check as a hard step before recruiter handoff
Bias auditing as an after-the-fact dashboard would have produced disparities that nobody acted on. Making bias-check a hard step before the recruiter saw the shortlist meant disparities were caught and the fairness summary was always current with the actual shortlist content.
Generated per-candidate rationale tied to specific resume content
Black-box scores would have failed recruiter trust and customer compliance simultaneously. Per-candidate rationale linked to specific resume content let recruiters validate the scoring quickly and gave compliance teams the explanation surface they needed for review.
Made fairness summaries shareable with customers’ legal teams
The fairness story had to extend beyond the platform to the customer’s own compliance posture. Generating disclosure-ready summaries per shortlist gave customers documentation they could share with their legal teams, which moved the conversation from "is this fair" to "can we represent it as fair," which the documentation answered.
What changed for the client
screening time
Reduction in recruiter time per requisition spent on initial screening across the rollout cohort over the first month.
time to hire
Reduction in median time from requisition open to candidate accepted across customer accounts in the first full quarter post-launch.
protected features used
Name, address, school prestige, and demographically correlated attributes excluded from the scoring path by design, audited continuously.
fairness summaries
Disclosure-ready summaries available per shortlist for customer compliance review and external representation if required.
The tools behind the system
Built with a deliberate stack chosen for production reliability and operational velocity.
Lessons learned from the build
Treating fairness as a design constraint rather than a checkbox made the rest of the system simpler. Excluding protected attributes by design produced a scoring path that did not require constant defense and freed the team to focus on accuracy and speed within that constraint.
Pre-handoff bias-check changed the operational shape of the system. After-the-fact dashboards would have produced disparities that lived as known issues. Making the check a hard step meant the shortlist that reached the recruiter was the shortlist that survived the audit, with no separate review backlog.
Per-candidate rationale earned recruiter trust faster than accuracy claims would have. Recruiters who saw the same signals they would have used themselves accepted the rankings and started overriding less. Rationale exposure is a trust multiplier, not a transparency line item.
Similar delivery work usually starts in these service areas
If you are exploring a similar product, workflow, or implementation challenge, these are the service tracks that usually fit best.
Where this project sits in the bigger market picture
Patterns for AI features, internal tooling, and product delivery in SaaS businesses.
Build a result-driven AI product with a team that has shipped before
If you are exploring a similar product, workflow, or AI use case, we can help scope the right architecture, delivery model, and first milestone.
Related case studies worth reviewing next
Have an AI idea, messy workflow, or product vision? Let's make it buildable.
Bring the problem. We'll help shape the product, define the architecture, and show the fastest path to a serious first version.
A practical first roadmap in the discovery call
Architecture, timeline, and delivery options in plain English
Security, scalability, and reliability discussed upfront
Model registry
softus-rag-v4.2
187ms
Latency
128k
Context
$0.004
Cost / req
Evaluation suite
Deploy pipeline
prod / canary 25% — healthy