Skip to main content
Xalicon

HRTech

An applicant tracking platform with explainable candidate matching

A recruitment platform where structured CV parsing and explainable matching cut screening time without removing recruiter judgement.

  • Next.js
  • TypeScript
  • Node.js
  • PostgreSQL
  • pgvector
  • Redis
  • AWS
Start a similar projectAll case studies

Sample case study — illustrative, not a delivered client engagement

Northwind Talent is an invented company used to demonstrate our case-study format and our approach to this kind of work. The technical detail reflects how we would genuinely deliver it; the client name is fictional and the figures under Results are design targets rather than measured outcomes.

Client

Northwind Talent

Illustrative HRTech scale-up, 40 staff, mid-market recruitment

Industry
HRTech
Year
2025

The business challenge

The team ran hiring pipelines for mid-market employers through a mix of a generic CRM, shared spreadsheets and email. A single role could attract several hundred applications, and recruiters were spending most of their day reading CVs to produce a shortlist. Screening quality varied by recruiter and by how late in the day an application arrived, and there was no defensible record of why one candidate progressed and another did not.

Objectives

  • Replace spreadsheet-based pipeline tracking with a single candidate record
  • Reduce the time between application and first recruiter decision
  • Make shortlisting consistent and reviewable across the team
  • Keep a human decision at every point that affects a candidate
  • Support GDPR obligations for consent, retention and deletion

Discovery and strategy

What we did before writing code

  • Shadowed four recruiters through a full screening cycle and measured handling time per application
  • Analysed a year of historical placements to understand which signals actually predicted a successful hire
  • Mapped every place candidate data was stored, including personal drives and email attachments
  • Ran a workshop on automated decision-making risk with the client’s legal adviser
  • Agreed that the system would rank and explain, never reject automatically

Solution

What we built

  • Unified candidate record

    One record per candidate holding applications, documents, communications, interview feedback and consent state, replacing four disconnected sources.

  • Structured CV parsing

    Documents are parsed into normalised skills, roles, durations and education, with per-field confidence. Anything below threshold is flagged for a recruiter to confirm rather than silently accepted.

  • Explainable matching

    Candidates are ranked against a role profile using semantic similarity over structured attributes. Every score shows the specific evidence behind it, and recruiters can adjust weightings per role.

  • Structured scorecards

    Interview feedback is captured against defined criteria rather than free text, making comparison across candidates and interviewers meaningful.

  • Retention automation

    Consent state and retention periods are enforced by scheduled jobs, with automated candidate notification and verified deletion.

  • Recruiter analytics

    Funnel conversion by stage, source and role, so the team can see where pipelines stall rather than guessing.

Architecture

How it fits together

Platform architecture

A conventional web application with a separate document-processing pipeline, so parsing load never affects interactive performance.

  1. Web applicationNext.js App Router, server-rendered recruiter workspace
  2. Application APINode.js service with role-based access control
  3. Document pipelineQueued parsing and extraction with retry and dead-letter handling
  4. Search and matchingPostgreSQL with pgvector for structured and semantic retrieval
  5. Consent and retentionScheduled enforcement jobs with verified deletion
  6. AnalyticsEvent stream feeding funnel and source reporting

Design approach

Decisions in the interface

  • Designed the recruiter workspace around the screening queue, since that is where most of the working day is spent
  • Every match score is expandable to show the evidence behind it, so recruiters can verify rather than trust
  • Bulk actions kept deliberately limited to prevent accidental mass rejection
  • Candidate-facing pages built to WCAG 2.2 AA and tested on low-end mobile devices
  • Consent and data-use language written with the client’s legal adviser rather than borrowed from a template

Development process

How the work ran

  • Two-week cycles with a working environment reviewed by recruiters every Friday
  • An evaluation set of 400 historical CVs with recruiter-verified extractions, used to score every parsing change
  • Matching quality measured against historical placement outcomes rather than perceived relevance
  • Playwright coverage of the screening, interview and offer journeys
  • Load testing against a synthetic dataset representing three years of application volume

Technology stack

What it runs on

  • Next.js
  • TypeScript
  • Node.js
  • PostgreSQL
  • pgvector
  • Redis
  • AWS

Challenges resolved

What went wrong, and what we did

Every project has these. A case study that omits them is a brochure.

  • Parsing quality across inconsistent CV formats

    Early extraction struggled with multi-column and scanned CVs. Adding a layout-aware pre-processing step and routing scanned documents through OCR before extraction raised field-level accuracy substantially on the evaluation set.

  • Avoiding automated adverse decisions

    Ranking risked becoming a de facto filter. The interface was changed so low-ranked candidates remain visible and require an explicit recruiter action, with the reason recorded.

  • Deletion across every store

    Candidate data existed in the database, object storage, the search index and analytics. A single deletion orchestrator now handles all four and produces a verification record.

Results

Objectives and design targets

Because this is a sample case study, these are the objectives and design targets agreed in scope — not measured business results.

  • Time to first decision

    Target: same day

    Illustrative target agreed during discovery, not a measured outcome

  • Screening consistency

    Structured scorecards on 100% of interviews

    Design target for the platform

  • Data sources consolidated

    4 systems into 1 record

    Scope defined during discovery

Client feedback

In their words

Placeholder quote

This quote is sample content shown to demonstrate the layout. It is not attributable to a real person or organisation and will be replaced only with a genuine, approved testimonial.

  • “Sample content for design demonstration. This quote is illustrative and is not attributable to a real client.”
    Sample Client ContactHead of Talent Operations (illustrative)

Related services

The practices behind this work

More work

Other case studies

All case studies

Similar situation?

Start a project like this one

Tell us where you are. We will tell you how we would approach it and what we would want to understand first.

Start Your ProjectTalk to us first

Prefer email? contact@xalicon.co