Name: ALEFE VITOR ALMEIDA GADIOLI
Publication date: 10/12/2025
Examining board:
Name |
Role |
|---|---|
| ALBERTO FERREIRA DE SOUZA | Coorientador |
| CLAUDINE SANTOS BADUE | Presidente |
| FRANCISCO DE ASSIS BOLDT | Examinador Externo |
| LUIS ANTONIO DE SOUZA JUNIOR | Examinador Interno |
Summary: The complexity of today’s financial markets demands solutions that combine continuous analysis,
effective risk management, and efficient communication. This dissertation presents PeterAI, a
multi-agent Artificial Intelligence framework that provides investment consultancy via WhatsApp,
combining Large Language Models (LLMs), chain-of-thought reasoning (CoT), and portfolio
optimization. The system is composed of specialized agents for each asset class, processes
multimodal inputs, and integrates real-time financial data with auditable explainability aligned
with regulatory requirements. For empirical validation, a controlled comparative study was
conducted over seven months (June–December 2024) with 120 real clients, stratified by risk
profile and financial objectives. Participants were equally divided between portfolios managed by
PeterAI and by specialized human advisors, each managing at least R$100,000. The results show
that PeterAI achieved statistically superior performance, with an average return of 8.29% ± 0.30%
compared to 2.57% ± 0.99% from human advisors, alongside lower volatility (1.45% ± 0.32%)
and a higher Sharpe Ratio (2.67 ± 0.85). The maximum drawdown was 46% lower than the control
group, and robustness tests confirmed superior performance even under market stress. Ablation
studies validated the importance of each component: removing the memory module increased
volatility by 46%, and omitting risk terms drastically reduced the Sharpe Ratio—confirming
the effectiveness of the integrated architecture. The main contributions include: (i) a scalable
MAS-LLM architecture for large-scale personalization; (ii) a robust error-handling pipeline
with ambiguity resolution; (iii) auditable explainability through CoT with immutable logs; and
(iv) empirical validation in a real-world environment. Identified limitations include regulatory
certification requirements and computational scalability challenges. The results suggest that the
proposed approach is feasible in a real-world setting and may contribute to more consistent
recommendations aligned with the risk profile, within the evaluated context and time period.
