Abstract
This paper introduces Asha Bot, an AI-powered conversational agent designed to address the unique challenges faced by women re-entering the workforce after career breaks. Through a combination of natural language processing, bias detection algorithms, and personalized recommendation systems, Asha Bot provides tailored career guidance, detects gender bias in job descriptions, and offers upskilling pathways. Our implementation combines the Rasa framework with custom bias detection modules and a hybrid recommendation system. Evaluation with participants demonstrated significant improvements in job application completion rates (+35%), interview confidence scores (+40%), and reduced time-to-employment compared to traditional career platforms. These results suggest that AI-powered conversational agents with domain-specific customization can effectively support women's career reintegration, addressing both practical and systemic barriers to workforce participation.
Introduction
Women's workforce participation continues to face significant challenges globally, with career breaks disproportionately affecting long-term career trajectories. Studies consistently show that women who take career breaks, particularly for caregiving responsibilities, face substantial obstacles when attempting to return to work. These challenges include skills depreciation, confidence issues, hiring biases, and lack of flexible opportunities. In India alone, 70% of women who take career breaks struggle to return to the workforce, with 69% facing bias during the hiring process.
Traditional job platforms and career services typically fail to address the unique needs of this demographic, focusing primarily on active job listings rather than the comprehensive support required for successful re-entry. This gap presented an opportunity for me to develop a technological solution providing personalized, accessible, and bias-aware support.
My project addresses three primary challenges faced by women returning to the workforce:
Skills and confidence gaps: Extended absences from professional environments often lead to outdated technical skills and diminished professional confidence
Structural and social barriers: Gender-biased hiring practices, inflexible work arrangements, and lack of re-entry programs create systemic obstacles
Insufficient personalized guidance: Generic career advice fails to address individual circumstances, skills, and career aspirations
My work makes the following contributions:
A comprehensive system architecture for career guidance using conversational AI
Novel techniques for bias detection and mitigation in career-related content
A hybrid recommendation approach tailored to career transitions
Empirical evidence on the effectiveness of AI support for workforce re-entry
Related work
AI in Career Guidance and Counseling
AI applications in career guidance have evolved from simple rule-based systems to sophisticated machine learning models. Recent systems like CareerLeaf and PathAdvisor demonstrate the potential of recommendation algorithms in suggesting career paths, though they typically lack the personalization required for career break transitions.
Gender Bias in AI and Employment
Research has consistently identified gender bias in natural language processing systems with particular implications for employment applications. Bias manifests in resume screening, job recommendation, and interview evaluation. My work builds on bias detection approaches from Dixon et al. and debiasing techniques from Bolukbasi et al. to create employment-specific bias mitigation.
Conversational Agents for Support and Guidance
Conversational agents have shown effectiveness in psychological support, educational guidance, and health coaching. However, few systems have specifically targeted career guidance, and fewer still address the unique challenges of women re-entering the workforce.
Methodology
System Architecture
I designed Asha Bot with a modular architecture comprising five key components:
Conversational AI Engine: Powered by Rasa framework for intent recognition, entity extraction, dialog management, and response generation
Bias Detection Module: Custom NLP pipeline leveraging Perspective API for identifying gender-biased language
Recommendation System: Hybrid approach combining collaborative filtering and content-based methods
User Profile Management: Dynamic user modeling based on conversation history and explicit inputs
Web/Mobile Interface: React-based frontend with voice input capabilities and responsive design
Conversational AI Implementation
I implemented the conversational component using Rasa Open Source, with custom actions written in Python. The training dataset comprised manually crafted domain-specific utterances, distinct intents relevant to career guidance, and entity types including skills, job roles, and industries.
The dialog policy was configured as a combination of TED policy for handling complex conversations, rule-based policy for specific workflows, and fallback policy with an appropriate threshold.
Bias Detection and Mitigation
My bias detection module consists of:
Pattern-based detection: A dictionary of common biased phrases in job descriptions
Perspective API integration: Custom configuration to detect subtle gender bias
Domain-specific classifier: A fine-tuned BERT model trained on manually labeled job descriptions.
For detected bias, I implemented three mitigation strategies:
Highlighting: Visually indicating potentially biased phrases
Explanation: Providing context on why the phrase may discourage applicants
Reframing: Suggesting alternative, gender-neutral phrasing
Recommendation System
The recommendation engine uses a hybrid approach:
Profile-based filtering: Matches user profiles to curated datasets of job listings, upskilling resources, and return-to-work programs
Collaborative component: Identifies patterns from similar users, weighted by background similarity, goal alignment, and career break duration
Temporal component: Accounts for recency of listings and skill demand trends
Experiments
Study Design
I conducted a mixed-methods evaluation with women who had experienced career breaks of 6 months or longer. Participants were recruited through women's professional networks in India and represented diverse industries, career stages, and break durations.
The study consisted of:
Randomized controlled trial: Participants were assigned to either the Asha Bot group or a control group using traditional job platforms
Usage period: Several weeks of system access
Data collection: Pre- and post-intervention surveys, system interaction logs, and semi-structured interviews
Evaluation Metrics
I measured effectiveness through:
Engagement metrics: Session frequency and duration, feature utilization, task completion rates
Career outcomes: Job application submission rate, interview attainment, job offer rate, time to employment
Psychological measures: Career self-efficacy (CSES scale), job search confidence (JSC scale), perceived bias encounters
System performance: Intent recognition accuracy, recommendation relevance, bias detection precision/recall
Results
User Engagement and Adoption
Asha Bot demonstrated strong engagement patterns with multiple sessions per user, substantial average session duration, and high completion rates for profile setup. The most frequently used features were skill assessment, job recommendation review, and bias detection for job descriptions.
Career Transition Outcomes
Compared to the control group, Asha Bot users showed significant improvements in key metrics:
Statistical analysis using independent t-tests confirmed these differences were significant with medium to large effect sizes.
Bias Detection Performance
My bias detection module achieved strong results in precision, recall, and F1-score:
Qualitative Insights
Thematic analysis of interview data revealed four key benefits of the system:
Confidence building: "The mock interview feedback helped me identify my strengths instead of focusing on the gap in my resume." (Participant)
Personalized guidance: "Unlike generic job sites, Asha understood my specific situation and suggested roles that actually matched my constraints." (Participant)
Bias awareness: "I didn't realize how many job descriptions subtly discouraged women from applying until Asha highlighted the patterns." (Participant)
Continuous support: "The consistent check-ins and progress tracking kept me motivated even after rejections." (Participant)
Discussion
Key Findings
My results demonstrate that an AI-powered career guidance system specifically designed for women re-entering the workforce can significantly improve career transition outcomes. The integration of bias detection, personalized recommendations, and continuous conversational support addresses multiple barriers simultaneously.
The substantial improvements in application completion rates and interview confidence suggest that addressing psychological barriers may be as important as providing practical job-matching services.
Limitations and Future Work
Several limitations should be acknowledged:
Sample demographics: My participant pool skewed toward technology and service sectors
Duration: The study period may not capture long-term career trajectory impacts
Regional context: The study was conducted primarily with participants from urban centers in India
For future development, I'm exploring:
Employer integration: Direct partnerships with employers offering return-to-work programs
Longitudinal studies: Tracking career progression over multiple years
Specialized modules: Industry-specific guidance for fields with unique challenges
Community features: Peer support and mentorship matching to complement algorithmic guidance.
Conclusion
Asha Bot demonstrates how AI-powered conversational agents can effectively address complex socioeconomic challenges when designed with attention to user needs, ethical considerations, and systemic barriers. By combining multiple AI technologies with domain-specific knowledge, I've created a system that significantly improves career transition outcomes for women re-entering the workforce.
The success of this project highlights the potential for AI systems to create meaningful social impact when developed to address specific needs of underserved populations. As I continue to refine and expand Asha Bot's capabilities, I hope this work inspires similar applications of AI for social good in other domains.
#Project Resources
GitHub Repository:
https://github.com/yessasvini23/Asha-Bot-Empowering-Women-in-Careers
Demo Video:
Refrences
Catalyst. (2023). Women in the workforce – Global: Quick take. https://www.catalyst.org/research/women-in-the-workforce-global/
Dastin, J. (2022). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters Technology News.
Hancock, A. B., & Rubin, B. A. (2023). Influence of communication partner's gender on language. Journal of Language and Social Psychology, 34(1), 46-64.
ILO (International Labour Organization). (2023). World employment and social outlook: Trends for women. https://www.ilo.org/global/research/global-reports/weso/trends-for-women2023/
Jhamb, D., Singal, A., & Samanta, P. K. (2022). Career breaks and women's employability in the Indian job market: Challenges and opportunities. International Journal of Human Resource Studies, 12(3), 234-251.
McKinsey & Company. (2024). Women in the workplace 2024. https://www.mckinsey.com/featured-insights/diversity-and-inclusion/women-in-the-workplace
Mishra, A., & Srinivasan, R. (2023). Second chance: Challenges and opportunities for women returning to work in India. Gender, Work & Organization, 30(1), 78-94.
WorldBank. (2023). Labor force participation rate, female (% of female population ages 15+) (modeled ILO estimate). https://data.worldbank.org/indicator/SL.TLF.CACT.FE.ZS
Acknowledgements
I would like to express my sincere gratitude to all those who supported this research project. First, I extend my thanks to the women who participated in our study, generously sharing their experiences and providing invaluable feedback that shaped Asha Bot's development. Their insights were crucial to creating a truly responsive system.
Special thanks to the women's professional networks across India who assisted with participant recruitment and provided platforms to discuss the unique challenges of workforce re-entry. I am also grateful to my academic advisors who offered guidance throughout the research process, particularly in developing rigorous evaluation methodologies.
This work was made possible through technical resources provided by the Rasa community, whose open-source framework formed the foundation of the conversational engine. I also acknowledge the contributions of the research teams behind the Perspective API and BERT, whose work enabled the bias detection capabilities central to this project.
Finally, I would like to thank my family for their unwavering support during the long hours of development and testing. This project represents a collaborative effort to address systemic barriers facing women in the workplace, and I am indebted to everyone who contributed to making it a reality.
Appendix
A. Example Dialogues
Below are representative dialogue examples demonstrating key features of Asha Bot:
Example 1: Initial Assessment
User: Hi, I'm looking to return to work after a 3-year break.
Asha: Welcome! I'm here to help you navigate your return to work. Could you share a bit about your previous experience and the field you're interested in?
User: I was in marketing for 5 years before taking a break for my children.
Asha: Thank you for sharing. Marketing has evolved significantly in recent years. Are you looking to return to a similar role, or are you considering exploring new areas within marketing?
User: I'd like something similar but with more flexible hours.
Asha: I understand flexibility is important to you. Marketing offers several remote and flexible opportunities, particularly in digital marketing and content creation. Would you like me to assess your current skills to identify potential gaps and strengths?
#Example 2: Bias Detection
User: I found this job that sounds perfect but I'm not sure if I qualify.
Asha: I'd be happy to review it with you. Please share the job description.
User: [Shares job description containing: "Looking for dynamic, aggressive go-getters who can work around the clock when needed."]
Asha: I've analyzed this job description and noticed some potentially biased language. Phrases like "aggressive go-getters" and "work around the clock" often discourage qualified women from applying. Research shows men apply when they meet ~60% of requirements, while women typically wait until they meet nearly all. Would you like suggestions for how to approach this opportunity?
Example 3: Upskilling Guidance
User: I think my technical skills are outdated. It's been 4 years since I worked in web development.
Asha: That's a common concern. The web development landscape does evolve quickly. Based on your profile and previous experience with JavaScript and React, I'd recommend focusing on these high-impact skills first:
- Modern JavaScript (ES6+) features
- React Hooks and functional components
- State management with Redux or Context API
Would you like me to recommend specific online courses that offer flexible learning schedules suitable for your availability?
B. System Components Details
Table B1: Conversational Model Performance
Table B2: Bias Detection Performance by Category
C. Study Participant Demographics
Table C1: Participant Demographics
