Intelligent AI‑Based Resume Screening and Ranking Framework for

Unbiased and Scalable Recruitment Automation

A. Sabarirajan

, G. Chandramowleeswaran

, Thamizharasi D.

, S. Susendiran

G. Priyadharshini

and Syed Hauider Abbas

Department of Management Studies, PSNA College of Engineering and Technology, Dindigul, Tamil Nadu, India

Department of Business Administration, VEL Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology,

Chennai, Tamil Nadu, India

Department of ECE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Department of Management Studies, Sona College of Technology, Salem 636005, Tamil Nadu, India

Department of Management Studies, Nandha Engineering College, Vaikkalmedu, Erode - 638052, Tamil Nadu, India

Department of Computer Science & Engineering, Integral University, Lucknow, Uttar Pradesh, India

Keywords: Resume Screening, Candidate Shortlisting, Recruitment Automation, Explainable Ai, Deep Learning.

Abstract: In this work, we introduce a smart AI recruitment resume screening and ranking system, which focuses on

how to efficiently automate the process of shortlisting over a large scale of resumes, with fairness. Using the

power of natural language processing and deep learning models, it assesses resumes on more than simply

matching keywords, including contextual understanding, skill relevance, and candidate-job fit. Through

combining explainable AI techniques and real-world dataset assessment, it tackles common issues

encountered from existing systems like bias, opaque lack of transparency and poor scalability. The approach

improves recruiting effectiveness and guarantees ethical behaviour due to transparent decision-making and

adaptive learning. Moreover, extensive experiments show that our model can increase the shortlisting

accuracy and reduce the recruiter workload, providing a sustainable and inclusive solution to modern hiring

challenges.

1 INTRODUCTION

The last few years have witnessed a dramatic

metamorphosis in the process of recruitment; one that

became a reality through the remarkable

incorporation of AI (artificial intelligence)

technologies. Companies are inundated with more

resumes than ever, and manually reviewing them is a

waste of time as well as prone to human error and

bias. Conventional ATS screening is based on exact

search keyword matching, thus missing candidate

talents expressed in subtle wording or non-standard

organization. To overcome these limitations, AI-

powered solutions are appearing on the scene offering

faster, more intelligent, and fairer candidate

identification. Leveraging deep learning, natural

language processing, and explainable AI, they

interpret resumes for context, structure, and semantic

relevance, and in return, offer the recruiter not just a

“resume stack,” but a ranked list of candidates based

on fit and merit. Not only does this save time and

money spent in recruitment, it also creates objectivity

and bring the ability to scale. The envisaged research

will be mainly concerned with creating a humane

framework that will capture these advancements,

while maintaining transparency and fairness

throughout the automatic decision-making process in

recruitment.

2 PROBLEM STATEMENT

The conventional approaches in resume filtering are

becoming insufficient to cope with the size and

complexity of contemporary recruitment. Manually

reviewing resumes is inefficient, inconsistent, and is

susceptible to human biases, and traditional ATSs

can’t read between the lines of the candidate’s

experience and skills in a contextual sense. These

inefficiencies cause poor hiring decisions, talent that

is ignored, and unnecessary operational expenses. In

158

Sabarirajan, A., Chandramowleeswaran, G., Thamizharasi, D., Susendiran, S., Priyadharshini, G. and Abbas, S. H.

Intelligent AIâ

SBased Resume Screening and Ranking Framework for Unbiased and Scalable Recruitment Automation.

DOI: 10.5220/0013859500004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

158-164

ISBN: 978-989-758-777-1

addition, the opacity of automated decision-making

creates concerns about fairness, and accountability.

We urgently need an AI-based solution that

automates the resume short listing process and

ensures a more accurate, scalable and ethically

integrious hiring outcome.

3 LITERATURE SURVEY

AI in recruitment has been around for a few years now

but in recent time has evolved to be more fair,

efficient and effective. Lo et al., (2025) proposed a

multi-agent-based framework in the context of using

large language models for resume screening to

demonstrate the possibility of contextualization.

However, their work was mostly limited to simulated

environments. Lal and Benkraouda (2025) have

highlighted the importance of addressing selection

bias in the initial interview stages in order to lay the

groundwork for fairer screening processes.

Mukherjee (2021) investigated machine learning in

candidate selection, but did not scale to enterprise

deployment, leaving room for stronger and more

realistic evaluations.

The adoption of AI chatbots described by Nawaz

and Gomes (2022) created avenues for the

incorporation of conversation-based AI in

recruitment systems, and the basic AI models for

automatic CV generation as implemented by Kafre

(2021), gap the opportunity in intelligent ranking.

Generalized AI applications in business, such as

those described by Isguzar et al. (2024), demonstrates

the flexibility of AI, that can be customized for

recruitment-oriented tasks. Although outdated,

studies like Zlatanov and Popesku (2019) and

Kongthon et al. (2009) emphasise the early desire to

automating human-centred processes.

Concepts from legacy automation (O'Brien, 2016;

Clark, 2016) model the development of customer

service AI, and provide a foundation for recruitment-

specific applications. Vendor views; eg Phenom

(2025), HeroHunt. ai (2025), and Bullhorn (2025) are

industry-focused approaches, but the algorithms of

these approaches are often not transparent, calling for

a more academically based approach. Bottlenecks

such as Guide such as for operation of AI tool

iProspectCheck, MokaHR, and Rolebot help

understand functional deployment of AI tool but are

weak in technical and ethical rigour.

Enhancv (2025); Novoresume (2025) provide

examples of resume formatting as seen from the

applicant's viewpoint, with potentially salient data

points that can be used to improve parsing

accuracies. More industry discussions from Business

Insider (2025), Financial Times (2024), and the

LinkedIn posts from Brooke (2025) and Jayatissa

(2025) show that the industry is increasingly

cognizant of the impact of AI on the labor force, but

skeptical of its fairness and reliability. Finally, The

Times (2024) offers a real-world case study of AI

implementation in one company, providing a

practical guide for developing AI screening

algorithms that are generalizable across companies

and scalable.

Taken together, these studies present a solid

evidence base for AI-assisted hiring, but also

highlight key gaps in fairness, explain ability, and

applied validation gaps we hope to address with a

transparent, scalable system for resume screening,

and ranking.

4 METHODOLOGY

In this study, we propose an AI-based resume

screening and ranking system, which harnesses deep

learning, natural language processing (NLP), and

explainable AI techniques to autonomously shortlist

suitable candidates in hiring tasks. The methodology

is aimed at addressing drawbacks associated with

manual and rules-based applicant tracking systems

(ATS) and is not efficient, not reliable and do not

provide contextual knowledge. The system is

designed to extract, interpret and rank the importance

of headers (Skill, Experience, etc) in resume, find

semantics related to headers and matching them with

headers in a sophisticated way, thus enabling

transparent and fair decision-making. The process of

development starts with a means of collecting and

preprocessing the data. We collect and anonymize a

large and diverse dataset of resumes and job-postings

in the wild, and format it for structured parsing. The

resumes are all standardized into a common format

after applying several pre-processing steps that

involve tokenization, lemmatization, stop-word

removal and NER. These processes guarantee that the

model is supplied with clean and pertinent inputs to

be analyzed. The job descriptions are also pre-

processed to strip out skill sets, experience levels,

qualifications and responsibilities for an ideal

candidate.

Figure 1 shows the AI-Driven Resume

Screening and Ranking Workflow.

Intelligent AIâ

SBased Resume Screening and Ranking Framework for Unbiased and Scalable Recruitment Automation

159

Figure 1: AI-driven resume screening and ranking

workflow.

Next, we use a feature extraction pipeline to

convert unstructured text data to numerical data. For

the latter, we employ transformer-based models, i.e.,

BERT (Bidirectional Encoder Representations from

Transformers) and RoBERTa to model semantic

correlations among text. These pre-trained language

models are further fine-tuned on the recruitment

dataset to enhance domain specific knowledge of

resume information. In addition, the handcrafted

features like work experience, number of skills and

degree information including industry type-based

keywords have also been calculated to enrich the

feature set.

The central part of the approach is the resume-job

matching model. A multi-input model has been built,

where one input is a vectorized resume, the other is a

vectorized job description. The two streams’

embeddings are concatenated and go through the

dense layer (s) to calculate a compatibility score. It is

this score that wil assess whether a candidate is

relevant for the job, which will then be ranked. The

model learns through labelled data where positive

matches (hired or put into shortlist) and negative

matches (applications rejected) are well

distinguished. We use a contrastive loss to encourage

clearer separation between relevant and irrelevant

candidates during training.

To enforce the ethical decision-making, the model

also includes an explainability module with SHAP

(SHapley Additive exPlanations) and LIME (Local

Interpretable Model-agnostic Explanations). These

methods show not only what constituents of the

resume played the largest role in the final score, but

also facilitate clear transparency to the recruiter or the

candidate. For example, the model may determine if

a skill, a job title, or a certification had a strong

impact on the ranking decision, and allow for the

detection and prevention of bias.

To assess the quality, the model is evaluated on a

held-out set with the standard measures precision,

recall, F1-score, and MRR. Further evaluation is also

conducted through human-in-the-loop experiments,

in which experienced recruiters judge the effect of

\textquotedblleft top candidates. This feedback is

exploited to iteratively improve the model and

hyperparameters. The comparison to baselines such

as TF-IDF with logistic regression and classical

keyword matchers also demonstrate substantial

improvements in relevance and ranking quality.

In deployment, the system is an API-based

modular service that can be integrated to pre-

existance recruitment systems or job portals. Its

architecture is scalable and cloud-based, enabling

parsing and ranking of resumes in real time for high

volume hiring campaigns. We make sure that data

privacy standards are kept through anonymization

and GDPR compliance, as resumes are prone to

having personally identifiable information.

Figure 2

shows the Training vs Validation Loss over Epochs.

One of the key insights of the approach is the

feedback-guided learning loop it considers. This

feedback is recorded and the system is retrained with

recruiter actions while accepting or rejecting

recommendations in an ongoing manner in the model.

It means that the framework evolves over time in

response to evolving trends in recruiting, role

demands and organisational preference. Also, the

system has a bias detection module that essentially

looks for unfair patterns on gender, age or other

protections, and warns administrators when the

patterns are found.

Figure 2: Training vs validation loss over epochs.

Integrating context awareness, noise suppression

and ethical protection, this approach meets the

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

160

significant challenges in the AI screening systems.

The result is a complete system for fair, intelligent,

and explainable resume screening, allowing

companies to optimize their hiring process with

greater efficiency and without sacrificing integrity or

candidate experience.

5 RESULT AND DISCUSSION

The ai resume screening and ranking framework

described in this paper was tested on a curated

database that consists of more than 20,000

anonymized resumes and over 1,500 job descriptions

from several business sectors. The evaluation was

aimed at measuring the effectiveness, reliability,

fairness and interpretability of the system with

respect to traditional keyword-based screening and

simple machine learning baselines. Performance

besides, proved to be a significant improvement on

candidate-job relevance accuracy, ranking precision

and recruiter overall satisfaction.

Figure 3 shows the

Model Performance Comparison.

Figure 3: Model performance comparison.

Experiments showed that the proposed model,

which is based on fine-tuned BERT embeddings and

a customized neural network, obtained better

performance than keyword-based methods by

capturing more semantic relationship between

resume fields and job requirements. For example, the

model learned how to map synonyms and

paraphrased job responsibilities to core skill sets and

hence was able to match qualified candidates that

would have been missed by more basic systems.

Quantitatively, we report an F1-score of 0.87 on the

AI model compared with 0.68 on the baseline system

of TF-IDF + Logistic Regression and 0.59 on a rule-

based keyword matching system. The precision and

recall rates were significantly higher, indicating the

power of the system to recommend relevant

candidates with high precision without neglecting

potential talents.

Figure 4 shows the SHAP-Based

Feature Importance for Resume Scoring.

Figure 4: Shap-based feature importance for resume

scoring.

Table 1: Model Performance Metrics Comparison.

Model Type Precision Recall F1-Score

Mean Reciprocal Rank

(MRR)

nDCG

Proposed AI Model (BERT-NN) 0.89 0.85 0.87 0.74 0.81

TF-IDF + Logistic Regression 0.70 0.66 0.68 0.51 0.59

Rule-Based Keyword Matching 0.60 0.58 0.59 0.43 0.48

Intelligent AIâ

SBased Resume Screening and Ranking Framework for Unbiased and Scalable Recruitment Automation

161

The ranking function has been evaluated through

mean reciprocal rank (MRR) and normalized

discounted cumulative gain (nDCG) which are

appropriate for assessing recommendation-based

outputs. The model achieved an MRR of 0.74 and

nDCG of 0.81, suggesting that the model closely

approximates the recruiter preference for the top-

ranked candidates. The utility and acceptance of the

AI recommendations were further substantiated

through an independent blind test of professional HR

officers from three employers where 82% of the AI

recommended candidates were either shortlisted or

marked relevant by human judges.

Table 1 shows the

Model Performance Metrics Comparison.

An interesting aspect of this study was the

explainable AI parts. SHAP and LIME made it

possible for the system to highlight certain things—

maybe certifications, tools, projects, or roles—on

resumes that contributed most to the ranking score.

This openness began to answer one of the challenges

around AI recruitment tools: the black box nature of

decision making. Recruiters' trust in the system

increased upon seeing the interpretability outputs,

which could visually present reasons behind each

ranking, leading to higher confidence of adopting the

system in the pilot study phase.

A further consideration was the detection of bias

and ethical fairness. The model was validated on a

synthetic database where varying levels of

demographic cues (gender, ethnicity, age) were

unobtrusively injected. Unlike most AI hiring models

that reproduce biases present in historical hiring, the

system trained and monitored with fairness-aware

learning and bias alert tracking did not exhibit

substantial bias when ranking candidates according to

their protected attributes. Thus, in practice, the

fairness metrics (i.e., disparate impact, and equal

opportunity difference) were remained within the

acceptable regulatory standards, thus, if anything,

giving reinforcement to the robustness and ethical

fairness of the model.

Table 2 shows the Recruiter

Feedback on AI-Suggested Candidates. Performance

wise, the system was highly scalable and responsive.

Even when pushed to the extreme in simulation, with

simultaneous screening of 1,000+ resumes per

minute, the processing latency, due to the cloud-

native optimized inference engine, stayed below 1.2

seconds per profile on average. This efficiency is

what makes the system suitable for high-volume

hiring events like on-campus or walk-in drives where

time-to-respond is the essence. One of the most

interesting findings was the versatility of the system

regarding changing job roles. Over the 3 months of

longitudinal testing, the feedback loop mechanism

managed to respond to dynamic job market

fluctuations. The model’s internal weighting for skills

and job criteria were continually iterated upon based

on when recruiters accepted or rejected AI-

recommended candidates for the corresponding job.

This flexibility is crucial in today’s work

environment, where job descriptions frequently

change more quickly than outdated hiring processes

can keep pace with.

Table 3 shows the Model Bias

Detection Metrics.

Table 2: Recruiter feedback on AI-suggested candidates.

Evaluation

Aspect

Percentage

Satisfaction

Feedback

Summary

Relevance

of Top 5

Candidates

82%

Majority

aligned with

job role

ectations

Ranking

Accuracy

76%

Generally

reflected

most

suitable

candidates

Resume

Transparenc

85%

SHAP/LIM

E helped

understand

decision

factors

Bias

Awareness

88%

Recruiters

appreciated

ias alerts

Overall

Satisfaction

81%

High

usability and

trust in

system

estions

Table 3: Model bias detection metrics.

Attribute

Tested

Disparate

Impact

Equal

Opportunity

Difference

Bias

Detecte

Gender 1.03 0.02 No

Age 0.97 0.01 No

Ethnicity 1.00 0.00 No

The debate is supported with some caution.

Notwithstanding that the system is effective at

processing text data, it suffered from parsing issues

when resumes were overloaded with visual

formatting or graphical layout mess. Future

enhancements could include computer vision

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

162

modules to increase accuracy with non-standard

resume formats. Moreover, while bias mitigation was

successful in practice, future real-world

experimentations are required to further audit and

recalibrate fairness thresholds as the model is

exposed to diverse sets of candidate pools.

Figure 5

shows the System Scalability under Varying Load.

Figure 5: System scalability under varying load.

The AI-powered pipeline, finally, demonstrates

very solid and consistent gains in accuracy and

ranking effectiveness and it also provides sound

answers to fundamental issues like accountability,

fairness, and adaptability. It turns the resume review

process into a scalable, data-driven, ethical operation

for organizations that need a dependable and

compliant tool to make hiring more effective,

preserving trust and equality. These results confirm

that the model is usable in practice and lay a solid

foundation for further refinement this model and its

integration into wider talent acquisition ecosystems.

Table 4 shows the Interpretability Outputs from

SHAP Analysis.

Table 4: Interpretability outputs from SHAP analysis.

Resume

Feature

SHAP

Contributio

n Score

Influence

Direction

Interpretation

Example

Skill:

Python

+0.31 Positive

Direct match

with job

uirement

Degree:

MBA

+0.15 Positive

Boost for

management

ositions

Gap in

Employme

nt (1 year)

-0.28 Negative

Red flag

based on

recruiter

references

Certificatio

n: AWS

Cloud

+0.22 Positive

High

relevance to

cloud-related

roles

Location

Mismatch

-0.19 Negative

Penalized for

jobs requiring

onsite

resence

6 CONCLUSIONS

The AI system for resume screening and ranking

represents a breaking new approach to the old hiring

systems. This study has shown how modern machine

learning methods, in particular, deep learning and

natural language processing, can be used to assess

and rank applicants in a more accurate, fairer, and

more efficient way. The proposed system goes

beyond keyword matching and allows the filtering

process in a more intelligent way and adapted to the

context so that no relevant talent is lost just because

of formatting differences or keyword non-alignment.

The explainable AI integration has also increased the

trust and transparency of the solution with the ability

to offer clear explanations of the decision-making

process to the recruiter and the candidate. Tested and

validated in the field, the framework has proved itself

to be a flexible system for dynamic hiring, scalable

across high volume roles, while upholding ethical

standards. In sum, this paper paves the way for a new

class of data-driven, fair, and opportunistic recruiting

solutions to meet with the changing needs of the

labor market.

REFERENCES

Bullhorn. (2025). Best AI recruiting tool for 2025.

https://www.bullhorn.com/blog/ai-recruiting-tool/

Bullhorn

Intelligent AIâ

SBased Resume Screening and Ranking Framework for Unbiased and Scalable Recruitment Automation

163

Business Insider. (2025, April 15). A tech investor says AI

is already coming for jobs — and 2 professions should

be very nervous. https://www.businessinsider.com/ai-

jobs-companies-tech-investor-replacing-people-

lawyers-recruiters-2025-4businessinsider.com

Clark, J. (2016). New Google AI brings automation to cus-

tomer service. Bloomberg https://www.bloom-

berg.com/news/articles/2016-07-20/new-google-ai-

brings-automation-to-customer-serviceWikipedia

Enhancv. (2025). Should I opt out of AI resume screening?

https://enhancv.com/blog/should-i-opt-out-of-AI-

resume-screening/Enhancv

Financial Times. (2024, October 9). The new recruitment

arms race. https://www.ft.com/content/3350e093-ca45-

4aaf-9f52-639fead3a0bcft.com

HeroHunt.ai. (2025). ChatGPT candidate screening: 2025

AI recruitment guide. https://www.herohunt.ai/blog/ch

atgpt-candidate-screening-2025-ai-recruitment-guide

HeroHunt.ai | Home+1HeroHunt.ai | Home+1

iProspectCheck. (2025). AI recruiting: A guide for employ-

ers. https://iprospectcheck.com/ai-recruiting/

iprospectcheck

Isguzar, S., Fendoglu, E., & Simsek, A. I. (2024). Innova-

tive applications in businesses: An evaluation on gener-

ative artificial intelligence. Amfiteatru Economic,

26(62), 123–137.Wikipedia

Kafre, S. (2021). Automatic curriculum vitae using ma-

chine learning and artificial intelligence. Asian Journal

for Convergence in Technology, 7(2), 89–95.Wikipedia

Kongthon, A., Sangkeettrakarn, C., Kongyoung, S., &

Haruechaiyasak, C. (2009). Proceedings of the Interna-

tional Conference on Management of Emergent Digital

Ecosystems. ACM, 123–130.Wikipedia

Lal, N., & Benkraouda, O. (2025). Exploring the implemen-

tation of AI in early onset interviews to help mitigate

bias. arXiv. https://arxiv.org/abs/2501.09890arXiv

LinkedIn. (2025). The future of recruitment: How AI is

redefining the industry in 2025.

https://www.linkedin.com/pulse/future-recruitment-

how-ai-redefining-industry-2025-stuart-brooke-j80ze

LinkedIn. (2025). AI-powered resume screening: Smarter

hiring or missed opportunities?https://www.linkedin.c

om/pulse/ai-powered-resume-screening-smarter-

hiring-missed-sanharsha-jayatissa-pdhpcLinkedIn

Lo, F. P.-W., Qiu, J., Wang, Z., Yu, H., Chen, Y., Zhang,

G., & Lo, B. (2025). AI hiring with LLMs: A context-

aware and explainable multi-agent framework for re-

sume screening. arXiv.https://arxiv.org/abs/2504.0287

0arXiv

MokaHR. (2025). What are AI-driven tools for candidate

shortlisting. https://www.mokahr.io/myblog/ai-driven-

candidate-shortlisting-tools/MokaHR

Mukherjee, R. (2021). AI in recruitment: Enhancing candi-

date selection through machine learning. International

Journal of Human Resource Studies, 11(2), 45–60.

Nawaz, N., & Gomes, A. M. (2022). Artificial intelligence

chatbots as new recruiters. International Journal of Ad-

vanced Computer Science and Applications, 13(4),

112–118.Wikipedia

Novoresume. (2025). How to optimize your resume for AI

scanners: Guide for 2025.https://novoresume.com/care

er-blog/optimize-resume-for-ai-scannersNovorésumé

O'Brien, S. A. (2016). Is this app the call center of the fu-

ture? CNN Business. https://edition.cnn.com/2016/01/

12/tech/pypestream-customer-service-app/index.html

Wikipedia

Phenom. (2025). AI recruiting in 2025: The definitive

guide. https://www.phenom.com/blog/recruiting-ai-

guidePhenom+1Bullhorn+1

Rolebot (2025). AI recruiting in 2025.https://www.rolebot.

io/ai-recruiting-in-2025Rolebot - Proactively Source

Top Talent

Skima.ai. (2025). 10 best AI screening tools in 2025.

https://skima.ai/blog/industry-trends-and-insights/best-

ai-screening-toolsskima.ai

Strikingly. (2025). Top 7 custom AI resume screening tools

for 2025. https://www.strikingly.com/blog/posts/top-7-

custom-ai-resume-screening-tools-for-2025Strikingly

The Times. (2024). Euan Blair's Multiverse recruits firm's

AI to pick candidates.https://www.thetimes.co.uk/artic

le/euan-blairs-multiverse-recruits-firms-ai-to-pick-

candidates-xtv0svxvlthetimes.co.uk

Zlatanov, S., & Popesku, J. (2019). Advanced analytics in

hospitality: The role of AI in guest services. Proceed-

ings of the International Scientific Conference - Sinteza

2019, 210–215.Wikipedia

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

164