AutoCare AI: An AI-Assisted Vehicle Service Management System

Abstract

Independent garages commonly manage repairs with paper job cards, verbal estimates and memory, so the accuracy and efficiency of their service cannot be measured. This paper presents AutoCare AI, a zero-cost web application built with Flutter and Firebase that digitises the garage workflow from walk-in to invoice, estimates a job\'s fault category, labour time and cost from the customer\'s free-text complaint using Google Gemini (with a keyword-rule engine for offline demonstration), suggests the least-busy suitable mechanic, predicts each vehicle\'s next periodic service and records manual, AI and actual outcomes on every job. The study follows a design-science approach with secondary, quantitative data: 31,236 owner complaints from the public NHTSA complaints database, labelled into seven of the application\'s fault categories and split into 24,988 training and 6,248 test records. A TF-IDF and Logistic Regression classifier reached 90.6 % accuracy (macro-F1 0.907), significantly above an 80 % target (exact binomial p ? 1.09 × 10?¹¹?), whereas the application\'s keyword rules reached only 57.6 %. The learned classifier was significantly more accurate than the rules (exact McNemar, 2,199 against 136 discordant pairs, p < 0.001). In a 250-day simulation, greedy assignment lowered the mean daily spread of mechanic workload to 1.49 h against 4.03 h for random assignment (Wilcoxon p ? 3.85 × 10??²). Gemini was not evaluated at scale because AI was kept switched off to guarantee zero cost; its few live estimates are reported only as anecdotal evidence.

Introduction

The study presents AutoCare AI, an AI-powered vehicle service management system designed to improve the efficiency, accuracy, and organisation of independent automobile garages, particularly small workshops in India. Traditional garages often rely on paper records, verbal estimates, and the owner's experience to diagnose faults, estimate repair costs, assign mechanics, and schedule future maintenance. These informal practices make it difficult to measure estimation accuracy, balance workloads, and ensure timely service reminders.

The primary objective of the study is to digitise the garage workflow and evaluate whether artificial intelligence can improve fault classification and mechanic assignment while operating at minimal cost. AutoCare AI is developed using Flutter for the web interface, Firebase Authentication and Cloud Firestore for data management, and Google Gemini for AI-based repair estimation. The system records customer and vehicle details, job cards, repair estimates, invoices, mechanic assignments, and actual repair outcomes. It also predicts future periodic servicing requirements and provides an AI Insights page to monitor estimation performance.

The literature review examines research on predictive vehicle maintenance, natural-language processing, machine learning-based fault classification, job scheduling, and large language models. It identifies a gap in affordable, data-driven service management systems designed specifically for independent garages, which generally lack the extensive historical records and resources available to large manufacturers and vehicle fleets.

The proposed system uses AI to classify customer complaints into fault categories, estimate labour hours and repair costs before GST, recommend suitable mechanics based on workload and specialisation, and predict upcoming maintenance. The owner reviews every AI-generated estimate before saving it, ensuring that human judgment remains part of the decision-making process. An offline keyword-based engine is available for demonstrations and testing, while the production system uses Google Gemini.

To evaluate fault classification, the study analyses public vehicle complaint data obtained from the US National Highway Traffic Safety Administration (NHTSA). Five methods are compared: a majority-class baseline, keyword-based rules, TF-IDF with Multinomial Naive Bayes, TF-IDF with Logistic Regression, and TF-IDF with a linear Support Vector Machine (SVM). The dataset contains 31,236 labelled complaints, with 6,248 records in the stratified test set.

The experimental results demonstrate that machine learning significantly outperforms the keyword-based approach:

• Logistic Regression: 90.6% test accuracy and 0.907 macro-F1 score.

• Linear SVM: 90.9% test accuracy and 0.910 macro-F1 score.

• Naive Bayes: 82.6% test accuracy and 0.807 macro-F1 score.

• Keyword-based rules: 57.6% test accuracy and 0.429 macro-F1 score.

• Majority-class baseline: 32.2% test accuracy.

Logistic Regression was selected for further use because its performance was close to that of the linear SVM, it provides class probabilities for confidence displays, and it showed stable cross-validation results. Statistical tests confirmed that the learned classifier exceeded the 80% accuracy target and significantly outperformed the keyword-based rules.

The study also investigates mechanic workload balancing through greedy assignment, although the supplied text does not include the detailed numerical results of that experiment. It acknowledges important limitations: the complaint dataset represents US vehicle owners who reported problems, rather than customers of Indian garages; the production Gemini model was not evaluated at scale; and the findings do not establish the real-world accuracy of repair-cost or labour-time estimates.

Conclusion

This study presented AutoCare AI, a zero-cost web application that digitises an independent garage\'s workflow, estimates fault category, time and cost from the customer\'s complaint, suggests mechanics, predicts the next service and records every estimate against the actual outcome. On 31,236 public NHTSA complaints, a learned classifier triaged faults with 90.6 % accuracy, significantly better than the application\'s keyword rules (57.6 %), and a 250-day simulation showed that greedy assignment spreads workload far more evenly than random assignment. The limitations are stated plainly: Gemini was not evaluated at scale because AI was kept switched off to guarantee zero cost; the data are U.S. defect complaints that differ from Indian garage complaints and lack two of the nine categories; time and cost accuracy could not be tested; the in-app sample data are synthetic; the scheduling result is simulated; and no user study was conducted. The next step is to evaluate the LLM and the time and cost estimates on real job cards.

References

[1] A. Theissler, J. Pérez-Velázquez, M. Kettelgerdes, and G. Elger, \"Predictive maintenance enabled by machine learning: Use cases and challenges in the automotive industry,\" Reliability Engineering & System Safety, vol. 215, art. no. 107864, Nov. 2021, doi: 10.1016/j.ress.2021.107864. [2] K. Kowsari, K. Jafari Meimandi, M. Heidarysafa, S. Mendu, L. Barnes, and D. Brown, \"Text classification algorithms: A survey,\" Information, vol. 10, no. 4, art. no. 150, 2019, doi: 10.3390/info10040150. [3] T. B. Brown et al., \"Language models are few-shot learners,\" in Advances in Neural Information Processing Systems 33 (NeurIPS 2020), 2020, pp. 1877–1901. [4] S. Mishra, L. Vassio, L. Cagliero, M. Mellia, E. Baralis, R. Loti, and L. Salvatori, \"Machine learning supported next-maintenance prediction for industrial vehicles,\" in Proc. Workshops of the EDBT/ICDT 2020 Joint Conference (DARLI-AP 2020), CEUR Workshop Proceedings, vol. 2578, Copenhagen, Denmark, 2020. [5] ?. Y?ld?r?m, A. D. Yücekaya, M. Hekimo?lu, M. Ucal, M. N. Aydin, and ?. Kalafat, \"AI-driven predictive maintenance for workforce and service optimization in the automotive sector,\" Applied Sciences, vol. 15, no. 11, art. no. 6282, 2025, doi: 10.3390/app15116282. [6] C. Zhang, W. Song, Z. Cao, J. Zhang, P. S. Tan, and C. Xu, \"Learning to dispatch for job shop scheduling via deep reinforcement learning,\" in Advances in Neural Information Processing Systems 33 (NeurIPS 2020), 2020, pp. 1621–1632. [7] J. Pavlopoulos, A. Romell, J. Curman, O. Steinert, T. Lindgren, and M. Borg, \"Automotive multilingual fault diagnosis,\" arXiv preprint arXiv:2210.06918, 2022. [8] F. Akhbardeh, T. Desell, and M. Zampieri, \"MaintNet: A collaborative open-source library for predictive maintenance language resources,\" in Proc. 28th Int. Conf. on Computational Linguistics: System Demonstrations (COLING 2020), Barcelona, Spain (Online), Dec. 2020, pp. 7–11, doi: 10.18653/v1/2020.coling-demos.2. [9] A. Dima, S. Lukens, M. Hodkiewicz, T. Sexton, and M. P. Brundage, \"Adapting natural language processing for technical text,\" Applied AI Letters, vol. 2, no. 3, art. no. e33, 2021, doi: 10.1002/ail2.33. [10] H. Ihshaish, P. Mayhew, T. M. A. Zayet, and A. Del Amo, \"The record is part of the task: Matched-record evaluation of text classifiers across maintenance, safety and recall reporting,\" arXiv preprint arXiv:2609.16267, 2026. [11] Z. Wang, Y. Pang, and Y. Lin, \"Large language models are zero-shot text classifiers,\" arXiv preprint arXiv:2312.01044, 2023. [12] M. Malyi, J. Shek, A. McDonald, and A. Biscaya, \"A comparative benchmark of large language models for labelling wind turbine maintenance logs,\" IET Renewable Power Generation, vol. 20, no. 1, art. no. e70212, 2026, doi: 10.1049/rpg2.70212. [13] S. Raschka, \"Model evaluation, model selection, and algorithm selection in machine learning,\" arXiv preprint arXiv:1811.12808, 2018 (v3, 2020). [14] J. Demšar, \"Statistical comparisons of classifiers over multiple data sets,\" Journal of Machine Learning Research, vol. 7, pp. 1–30, 2006. [15] A. R. Hevner, S. T. March, J. Park, and S. Ram, \"Design science in information systems research,\" MIS Quarterly, vol. 28, no. 1, pp. 75–105, 2004. [16] National Highway Traffic Safety Administration (NHTSA), \"NHTSA Datasets and APIs — Complaints,\" U.S. Department of Transportation, 2026. [Online]. Available: https://www.nhtsa.gov/nhtsa-datasets-and-apis [17] F. Pedregosa et al., \"Scikit-learn: Machine learning in Python,\" Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [18] W. G. Cochran, Sampling Techniques, 3rd ed. New York, NY, USA: John Wiley & Sons, 1977.

Copyright

Copyright © 2026 Om Singh, Perseus Bhavnagri. 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.