Main Menu
Participant Journey

How Two NUS MSBA Alumni Are Making AI-Assisted Hiring Explainable at X0PA AI

At X0PA AI, NUS MSBA graduates Ural Malik and Yu Zhe build the models behind hiring decisions at national scale. Between them they have shipped an in-house CV parser, explainable candidate-job matching, an asynchronous AI interview product, and a retrieval layer that lets assessors question a candidate's full profile in plain language. The common thread is not accuracy alone, but making every output answerable to the recruiter who has to act on it.

The challenge

Why more data did not make hiring easier

When every applicant clears the keyword filter, the filter has stopped working.

Recruitment now generates more data per candidate than any team can read. A single role at a large employer can attract thousands of applications, each with a CV, an application form, and increasingly a set of structured assessment responses. The constraint was never the volume of information. It was that none of it was comparable.

The first challenge is structure. CVs arrive in formats no two of which agree: multi-column layouts, tables inside tables, scanned PDFs, dates written six different ways, job titles that describe the same work in unrelated language. Before any model can rank a candidate, something has to turn that into structured, comparable data. Get that step wrong and every downstream score inherits the error.

The second constraint was that a CV could never tell the whole story. Skills, education and experience can show whether someone appears qualified, but they say less about how a candidate works, responds to situations or may fit with a team. That meant looking beyond the résumé — through behavioural signals, structured assessments and interview responses — to give recruiters a broader view of candidate suitability.

The third constraint was transparency. A model that ranks candidates well is not automatically a model a recruiter can use. Hiring is a decision someone has to justify, to a hiring manager, to a candidate who was not selected, and increasingly to a regulator. A score without a reason is not usable evidence, however accurate it is.

We realised quite early that accuracy alone was not enough. In recruitment, if a model gives a score but the recruiter cannot understand what is driving it, it is difficult for that score to support a real hiring decision.

— Yu Zhe, Senior Data Scientist, X0PA AI

What they built

Four systems, one requirement: show the reasoning

A parser, a matcher, an assessment product and a retrieval layer, each built so a recruiter can see why.

X0PA AI’s technology spans the recruitment journey from sourcing through to decision support. Four components carry most of the analytical weight.

The CV parser. Built in-house rather than licensed, it converts inconsistent résumé formats into structured candidate records. This is the foundation the rest of the stack depends on, and the reason it was brought in-house: an external parser’s failures are invisible and unfixable, and they propagate silently into every score built on top.

Candidate-job matching. Rather than keyword overlap, the matching models weigh skills, experience, education and job function together, surfacing candidates whose relevance a text search would miss. Explainability was a design constraint from the start, not a feature added afterwards. Persona Match extends this to fit beyond the strictly technical.

X0PA Room. The company’s asynchronous interview and assessment product, extending evaluation beyond the CV through structured video, audio and text-based responses. AI-assisted scoring gives recruiters a consistent basis for comparison across a large applicant pool, which manual review at that volume cannot provide.

Persona Match. Technical fit is only one part of candidate suitability. X0PA Persona assesses work styles, communication patterns and behavioural traits, while Persona Match looks at compatibility with the role, future manager or team. It can also provide broader team-level insights into how different working styles may interact, giving recruiters another perspective beyond qualifications and experience..

Generative reports and retrieval. In a recent public sector engagement, generative AI candidate reports were built to surface the reasoning behind an assessment rather than only its conclusion, paired with a retrieval-augmented layer that lets assessors query a candidate’s full profile in plain language. Work that meant reading detailed reports one by one became a matter of asking a question.

Extending AI beyond the platform. They also works with partners to bring its AI capabilities into their own systems and recruitment workflows. One example is its work with Workforce Singapore on MyCareersFuture, where X0PA has supported areas such as candidate recommendation, application scoring, skills taxonomy and suggested skills through customised AI services and APIs.

Assessors weren’t short of information about candidates. They were short of time to read it. The question was how to make the detail reachable without asking anyone to read all of it.

— Ural Malik, NUS MSBA Alumnus & Senior Data Scientist, X0PA AI

Key takeaways

What the Work Has Taught Us

01

The foundation sets the ceiling

Parsing and data structuring may be less visible than the models built on top, but errors at this stage can affect every downstream application.

02

Accuracy and usability are different objectives

A strong model is only valuable if recruiters can use and understand its output. Optimising for the real decision matters as much as optimising the metric.

03

Explainability has to be designed in

Useful explanations work best when considered during model design, rather than added only after a prediction is made.

04

Scale exposes robustness

Aggregate performance can hide uneven behaviour across different roles, industries and career backgrounds. Larger and more diverse deployments make robustness and consistency increasingly important.

05

AI supports the decision; it does not make it

The role of the system is to surface relevant information and reduce what recruiters need to review, while keeping human judgement at the centre.

06

Responsible AI is part of development

Fairness, robustness, explainability and data protection need to be considered throughout model development and evaluation, supported by frameworks such as AI Verify and strong data-governance practices.

07

Stay open to new ways of solving problems

Data science continues to evolve beyond traditional modelling. Generative AI, RAG and new search approaches can complement existing methods when they create practical value.

Ural Malik

Ural Malik worked in procurement analytics at The Smart Cube for five years before moving into recruitment AI, building cost models for clients across consumer goods, chemicals and manufacturing. That work depended on being able to defend every input when a supplier questioned the number, a habit that carried over. He graduated from the NUS Master of Science in Business Analytics in 2021, working during the programme as a student researcher under Professor Jussi Keppo and then at the SIA-NUS Digital Aviation Corporate Lab on reinforcement learning for pilot simulation training.

He joined X0PA AI in November 2021 as a Senior Data Scientist. He works on X0PA Room, the company’s asynchronous AI interview assessment product, and built the in-house CV parser that handles the inconsistent formats résumés arrive in, along with recommendation modules matching candidates to internship roles at partner institutions. In a recent public sector engagement, he built generative AI candidate reports designed to surface the reasoning behind an assessment rather than only its conclusion, alongside a retrieval-augmented layer that lets assessors query a candidate’s full profile in plain language.

Connect with Ural on LinkedIn ↗

Yu Zhe

Yu Zhe graduated from the NUS Master of Science in Business Analytics in 2021, following an earlier career in offshore engineering that developed her analytical thinking, structured problem-solving and modelling skills. The MSBA programme supported her transition into data science through applied training in machine learning, analytics and data-driven decision-making.

Since 2021, she has been a Senior Data Scientist and Data Protection Officer at X0PA AI, working on explainable candidate-job matching, talent recommendation, Persona Match and RAG-based candidate insights. Her work also includes supporting large-scale AI recruitment services for clients, with a focus on robustness, explainability and consistent performance. Alongside model development, she contributes to responsible AI by integrating the AI Verify framework into development practices and leading X0PA’s certification under Singapore’s SS 714 data protection standard.

Connect with Ural on LinkedIn ↗

What comes next

As CVs and applications become easier to generate and refine with AI, recruiters will need to look beyond credentials and self-reported information towards demonstrated skills, structured assessment and richer candidate insight. Expectations around transparency, fairness and accountability in AI-assisted hiring continue to rise alongside.

For Ural and Yuzhe, the next opportunity lies in extending systems across more of the recruitment journey while keeping their reasoning visible, and in developing matching approaches that surface capable candidates whose backgrounds do not follow conventional career or credential pathways.

Original Publish Date

Related articles

Participant Journey

From the Classroom to the CRM: How a Non-Technical CEO Built AI Tools That Run His Company Every Week

When Erick Villalobos, CEO of Singapore's Evermove Engineering, attended the NUS Business School Executive Education course AI for Talent Management and Organizational Design in early 2025, he was 39 years old, ran a 52-year-old company, and had never written a line of code. Six months later, his company runs a fully custom AI-built CRM, with no developers, no budget, and no technical team. This is the story of how a strategic mindset, not technical skill, made the difference.
Read more