Adaptive Multi-Agent AI System for Synthetic Data Generation and Consumer Persona Simulation: Architecture, Implementation, and Empirical Evaluation

Author

Prof. U.A.S. Gani, Saurabh Sharma, Sarthak Bankar, Shree Larokar, Aditya Tiwari


Abstract

 This paper presents the design, architectural implementation, and empirical evaluation of a unified Two-Agent Artificial Intelligence System addressing two core operational challenges in data science and product marketing: (1) generating realistic, statistically coherent tabular datasets without requiring exposure to sensitive authentic records, and (2) simulating pre-launch consumer psychographic reactions and purchasing intent across demographically diverse cohorts without expensive focus groups. The system implements two cooperating agents orchestrated under a unified web interface: Agent 1 (Synthetic Data Generation Agent), an 8-stage description-driven tabular synthesis pipeline that decouples schema planning from row sampling via large language models and parametric statistical distributions; and Agent 2 (Consumer Persona Simulation Agent), a 7-stage cognitive simulation pipeline that injects first-person psychographic conditioning across five benchmark personas, performs population-level simulation over N = 1,000 synthetic Indian consumers, and synthesizes price-conditioned customer reviews. Both agents share a fault-tolerant Gemini API fallback architecture (gemini-3.1-flash-lite -> gemini-3.5-flash-lite -> gemini-3.6-flash -> deterministic fallback). Extensive empirical evaluations demonstrate that Agent 1 generates datasets with preserved inter-field correlations and competitive downstream ML utility (ROC-AUC 0.75-0.79), while Agent 2 accurately differentiates product value propositions and decisively detects credibility collapses that generic LLM prompting fails to identify.


Keywords

Synthetic data generation, consumer persona simulation, multi-agent AI, LLM evaluation, marketing intelligence, adaptive pipeline, Gemini API, machine learning



Full Text:

Download Paper PDF


References


  1. Nadas et al., 'Synthetic Data Generation Using Large Language Models: Advances in Text and Code,' arXiv preprint, 2025.
  2. Various Authors, 'On LLMs-Driven Synthetic Data Generation, Curation, and Evaluation: A Survey,' arXiv preprint, 2024.
  3. M. H. Nahid and S. B. Hasan, 'SafeSynthDP: Leveraging LLMs for Privacy-Preserving Synthetic Data Generation Using Differential Privacy,' arXiv preprint, 2024.
  4. J. Argyle et al., 'Out of One, Many: Using Language Models to Simulate Human Samples,' Political Analysis, vol. 31, no. 3, pp. 337-351, 2023.
  5. S. Park et al., 'Generative Agents: Interactive Simulacra of Human Behavior,' in Proc. ACM UIST, San Francisco, CA, 2023, pp. 1-22.
  6. Google DeepMind, 'Gemini: A Family of Highly Capable Multimodal Models,' arXiv:2312.11805, Dec. 2023.
  7. Agrawal and P. Mittal, 'Consumer Behavior in Indian E-commerce: Price Sensitivity and Trust Dynamics,' Journal of Retailing and Consumer Services, vol. 67, 2022.

 

Share your valuable work from Social Media Buttons