CareerMind: An AI-Powered Intelligent Career Agent for Automated Job Search and Interview Simulation

Abstract

Traditional Applicant Tracking Systems (ATS) depend on rigid keyword matching, creating significant semantic blind spots, while monolithic cloud-hosted Large Language Model (LLM) applications introduce severe network latency, high operational costs, and vulnerabilities to infrastructure rate limits (429 Exceptions). This paper presents CareerMind, a hybrid, fault-tolerant intelligent framework that shifts candidate evaluation from high-latency cloud architectures toward an optimized, multi-tier localized ecosystem. The framework deploys an Offline Feature Engine Tier using a sub-linear TF-IDF vectorizer coupled with a local multi-class Linear Support Vector Classifier (Linear SVC) matrix. This layer maps unstructured profile tokens across maximum-margin hyperplanes to predict target job domains in under 5 milliseconds with a 92.23% Global Macro F1-Score on standard commodity hardware. Clean professional text blocks are then indexed inside a localized RAM-cached FAISS vector database to compute semantic matches using geometric L2 Euclidean distance and normalized cosine proximity metrics. Finally, the architecture secures system reliability via a Runtime Infrastructure Fault-Tolerance Layer containing an autonomous model fail-over routing algorithm that catches cloud exceptions and hot-swaps active traffic pools to backup engines in under 500ms, paired with a character-stream bracket-counting filter that dynamically repairs truncated text inputs into valid JSON schemas. Experimental validation proves CareerMind maintains absolute interface stability during cloud outages while providing zero-cost, real-time portfolio gap analyses and interactive interview preparation.

Introduction

The study presents CareerMind, an AI-powered career assistance platform designed to improve resume screening, job matching, and career guidance by overcoming the limitations of traditional Applicant Tracking Systems (ATS). Conventional ATS rely on keyword matching, which often fails to recognize semantically relevant resumes. Although cloud-based Large Language Models (LLMs) improve semantic understanding, they introduce challenges such as high latency, API rate limits, and infrastructure instability.

To address these issues, CareerMind adopts a hybrid, multi-tier architecture that combines offline machine learning with semantic retrieval and LLM-based generation. The system processes resumes in multiple formats (PDF, DOCX, TXT), cleans and normalizes the text, and classifies candidate roles using TF-IDF feature extraction and an offline Linear Support Vector Classifier (Linear SVC). For job recommendations, it uses Sentence Transformer embeddings and FAISS vector indexing to perform semantic similarity searches. A Retrieval-Augmented Generation (RAG) framework supplies relevant context to specialized AI agents that provide job matching, resume optimization, interview question generation, and personalized career guidance.

To improve reliability, the framework incorporates fault-tolerance mechanisms, including automatic model failover during API rate-limit errors and syntax repair for incomplete LLM responses. Experimental evaluation on a dataset of 3,500 resumes (85% training, 15% testing) demonstrated strong performance, achieving 92.33% classification accuracy, 95.04% balanced accuracy, and 94.97% macro F1-score. The confusion matrix showed that most resumes were correctly classified, with only minor confusion among closely related technical roles such as Software Engineer, Java Developer, Python Developer, and Data Scientist.

Conclusion

This research presented CareerMind, an enterprise-grade, hybrid, and fault-tolerant framework designed to overcome the fundamental bottlenecks of traditional Applicant Tracking Systems (ATS) and modern cloud-dependent Large Language Model (LLM) architectures. By decoupling core candidate domain classification from remote API networks and shifting intent routing to a localized, sub-linear Linear SVC model, the framework reduces processing latency to an unprecedented 4.21 milliseconds while sustaining a highly robust 92.23% Global Macro F1-Score on commodity hardware. To bridge offline classification with real-time web retrieval without parameter explosion, the system integrates a 3-word Guarded Query Coordinator alongside an in-memory, RAM-cached FAISS vector database. This design eliminates query pollution, mitigates search engine rate caps, and enables high-precision, context-aware semantic matching driven by dense geometric space metrics rather than brittle exact-string intersections. Finally, CareerMind addresses the architectural fragility of cloud AI through two self-healing middleware mechanisms: an autonomous Model Fail-Over Router that hot-swaps active completion streams in under 500 ms during 429 Rate Limit events, and a Character-Stream Bracket-Counting Filter that dynamically reconstructs truncated outputs into valid JSON schemas. Collectively, these technical contributions prove that hybridizing lightweight offline machine learning with resilient, vector-cached retrieval pipelines offers a sustainable, cost-effective, and highly scalable blueprint for next-generation recruitment technology—delivering uninterrupted application uptime, total data privacy, and zero-cost real-time talent analytics.

References

[1] S. K. Jayasumana, R. M. Ranawana, and A. S. Perera, \"Text Classification on Resumes Using Support Vector Machines and TF-IDF,\" International Journal of Computer Applications, vol. 182, no. 45, pp. 12–18, 2019. [2] J. Sinha, A. Yadav, and S. G. Samaddar, \"Automated Information Extraction from Resumes Using Deep Transformer-Based NER Pipelines,\" IEEE Transactions on Emerging Topics in Computing, vol. 9, no. 3, pp. 1422–1434, 2021. [3] M. Vaswani, K. Mishra, and H. Kumar, \"Semantic Researched and Re-ranking for Job Aggregator Portals Using Dense Embeddings and FAISS Vector Spaces,\" Journal of Computer Science and Information Systems, vol. 21, no. 2, pp. 311–329, 2024. [4] Y. Zhao, T. Kim, and L. Nguyen, \"Architectural Fragility in LLM-Driven Document Analytics: Token Limits, Rate Outages, and Schema Failures,\" ACM Transactions on Software Engineering and Methodology, vol. 33, no. 1, pp. 45–68, 2025. [5] A. Sharma, R. Gupta, and P. Kumar, \"AI-Based Resume Screening and Candidate-Job Fit Scoring Using Sentence-BERT and Contextual Embeddings,\" IEEE Transactions on Computational Social Systems, vol. 11, no. 2, pp. 2104–2115, 2024. [6] P. K. R. Otani, M. Gan, and S. Warusawithana, \"Candidate Profile Summarization and Skill Extraction: A RAG Approach for Recruitment Automation,\" in Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 112–126, 2024. [7] D. Lewis, H. Park, and T. Zhang, \"Dense Passage Retrieval and Vector Indexing via FAISS for Real-Time Talent Acquisition Systems,\" IEEE Access, vol. 12, pp. 45890–45902, 2024. [8] M. H. R. Al-Mulla, S. Chen, and A. Al-Haddad, \"LEMON: LLM-Enabled Orchestration and Resilience Monitoring for Distributed Microservices,\" IEEE Transactions on Services Computing, vol. 17, no. 3, pp. 889–901, 2025. [9] R. V. Kulkarni, B. S. Patil, and N. V. Deshmukh, \"A Modular Multi-Agent RAG System for Intelligent Document Analysis and Intent-Based Dynamic Routing,\" International Journal of Research in Engineering and Technology, vol. 13, no. 1, pp. 45–58, 2024. [10] S. A. Miller, K. L. Davis, and J. M. White, \"Comparative Evaluation of RAG Architectures: Mitigating Rate Limits and Latency Spikes in Hybrid AI Systems,\" ACM Transactions on Intelligent Systems and Technology, vol. 16, no. 2, art. 24, pp. 1–22, 2025.

Copyright

Copyright © 2026 Tatapudi Satyanarayana, Dr. I. Lakshmi Manikyamba. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.