Sugran: The Maharashtrian Recipe Generator — An AI-Powered Serverless Multimodal Web Application for Regional Culinary Preservation

A. Conclusion The successful implementation of Sugran validates the transformative potential of Large Multimodal Models when applied to hyper-local cultural preservation tasks. Three primary engineering conclusions have emerged: (1) Superiority of Client-Side Privacy Architecture: By adopting a decentralised, serverless architecture, Sugran proves that AI applications can be simultaneously powerful and privacy-preserving, establishing a replicable standard for privacy-first AI application design. (2) Prompt Engineering as a Reliable Algorithmic Discipline: Strict schema enforcement through persona-based meta-prompts can reliably transform conversational AI into a precise, deterministic data-generation engine. The JSON Forcing methodology achieved 100% parse success and is highly generalisable to other cultural documentation tasks. (3) Democratisation of Specialised AI: Sugran conclusively demonstrates that undergraduate engineering teams working with limited resources and no proprietary datasets can leverage general-purpose LMMs to outperform commercial CNN-based alternatives by significant accuracy margins (94% vs. 84.2%) in specialised cultural domains.

[1] Bhombe, P., Bharambe, N., Baviskar, S., Khatal, A., & Hemade, R. (2025). A Comprehensive Survey on AI-Based Maharashtrian Recipe Image Recognition and Regional Culinary Personalisation Using Multimodal Deep Learning. Dept. of CSE, NCER, Pune (Under Review). [2] Bossard, L., Guillaumin, M., & Van Gool, L. (2014). Food-101 - Mining Discriminative Components with Random Forests. Proceedings of ECCV, Zurich, 446-461. [3] Chen, J., Zhu, D., Li, X., Wang, L., & Liu, Y. (2023). A Comprehensive Survey of Large Multimodal Models: Architecture, Training Paradigms, and Applications. arXiv preprint arXiv:2306.13549. [4] Deshpande, M. (2022). Deep Learning for Maharashtrian Cuisine Recognition Using InceptionV3 Architecture. Journal of Emerging Technologies in Engineering Applications, 5(2), 34-41. [5] Gomez, L., & Perez, R. (2022). Accessibility Benefits of the Web Speech API in Culinary Mobile Applications. Proceedings of ACM CHI, New Orleans, LA, 1-12. [6] Google AI for Developers. (2024). Gemini 1.5 Flash Model Documentation and API Reference. Google LLC. https://ai.google.dev/docs [7] Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large Language Models are Zero-Shot Reasoners. NeurIPS, 35, 22199-22213. [8] Lo, F., Wang, X., & Huang, M. (2025). Nutritional Analysis from Food Images Using Large Multimodal Models. Proceedings of ACM Multimedia. [9] Patel, A., Joshi, R., & Kulkarni, S. (2025). Multilingual LLMs for Indic Languages: Marathi Benchmark Evaluation. International Journal of NLP Research, 12(1), 45-62. [10] Patel, J., Shah, V., & Mehta, N. (2023). MobileNetV3-Based Lightweight Indian Food Recognition. Proceedings of MISP, 441-449.