[{"jcr:title":"Computer Engineering students develop generative AI agent","cq:tags_0":"area-de-conhecimento:ciência-da-computação","cq:tags_1":"area-de-conhecimento:tecnologia","cq:tags_2":"programas:graduacao"},{"richText":"The Capstone project utilized recent computational technologies to create an autonomous shopping environment virtual system","authorDate":"19/08/2025 12h11","author":"Leandro Steiw","madeBy":"Por","tag":"area-de-conhecimento:ciência-da-computação","title":"Computer Engineering students develop generative AI Agent","variant":"image"},{"jcr:title":"transparente - turquesa - vermelho"},{"themeName":"transparente - turquesa - vermelho"},{"containerType":"containerTwo"},{"jcr:title":"Grid Container Section","layout":"responsiveGrid"},{"jcr:title":"Caio Ortega Bôa, Gustavo Eliziario Stevenson de Oliveira, João Pedro Rodrigues dos Santos, and Matheus Raffaelle Nery Castellucci","alt":"Caio Ortega Bôa, Gustavo Eliziario Stevenson de Oliveira, João Pedro Rodrigues dos Santos, and Matheus Raffaelle Nery Castellucci"},{"text":"In the first semester of 2025, Insper’s Capstone project proposed a relatively complex challenge centered on recent artificial intelligence (AI) technologies to Computer Engineering students. Caio Ortega Bôa, Gustavo Eliziario Stevenson de Oliveira, João Pedro Rodrigues dos Santos, and Matheus Raffaelle Nery Castellucci, all from the eighth semester, were tasked with developing the project “Negotiation between Generative AI Agents for Communication in Spatial Web,” under the guidance of professor Luciano Pereira Soares, for partner company NTT DATA.    The group worked on a system based on large language models (LLM), which are AI resources used in tools for text, image, and video generation, such as ChatGPT, Gemini, Llama, and DeepSeek. The interactive agent operates in a graphic environment similar to a game. “As these technologies are still in the consolidation phase, the project demanded a great research effort, exploring various alternatives until identifying the most suitable one for the project,” says Soares, academic coordinator of the Capstone program and Insper's Virtual Reality and Digital Games Lab.   In the report, the students explain that autonomous generative AI agents are computational systems with the ability for perception, inference, and action. They can make complex decisions that have the potential to transform areas as diverse as digital interaction in immersive environments, process automation, and the negotiation of goods and services. Faced with the challenge from NTT DATA, the Capstone demanded the creation of a generative AI agent system that performed user-oriented tasks in a virtual shopping environment.   To enable realistic autonomous interactions, the project combined generative AI and computational memory. The group found, during the tests of the first versions of the application, that memory was essential to make interactions between user and agent faster and more coherent —affecting the cost of operating the virtual environment. Largely, the AI didn’t register the sequence of the dialogue, so it tended to keep repeating the same question to the consumer. The solution was to create a local cache system that recorded previous interactions. The memory feature reduced interaction time from five minutes to a few seconds.   According to Soares, because it is an innovation project, with many uncertainties involved, there is a significant risk of not achieving a functional outcome at the end of the work. “Still, the students managed to present a complete prototype, and I understand that NTT DATA was satisfied with the result,” says the professor. “The group dedicated themselves a lot, and I noticed they invested considerable time to complete the project. Working in a team is always a challenge, but in the end, the group managed to organize themselves properly and deliver the project successfully.”   It was the fifth time NTT DATA participated in an Insper Capstone —previously called Engineering Final Project. Soares highlights the attention the company always dedicates to the students, closely following and helping to refine each project. Cauê Dias, Head of Digital Technology Innovation at the company in Brazil, reinforces the validity of the partnership. “We’ve really liked the work style we’ve developed and what Insper proposes,” Dias said. “NTT DATA focuses on seeking talents. More than developing solutions for a client's problems, we want the students to foster the innovation ecosystem, whether with a new technology or researching a topic the market has never seen, so that it also leaves a mark on their curriculum. I think this is the right term: we like the readiness of Insper students.”   Dias adds: “It’s challenging to find a digital research and development lab in Brazil like NTT DATA's. So the students are very curious about new concepts and understand easily what is being proposed because they come with enormous grounding from Insper. They have all the technical tools to quickly understand business cases. It becomes even enjoyable to discuss with young people who already have such technical knowledge. And the professors also provide great support.”     Coherent Interaction   Among other results, the students report that the agents generated well-structured dialogues, maintaining discourse cohesion and not forgetting previously discussed information. They also demonstrated natural and organic behavior, responding coherently to the conversation context. In João Pedro dos Santos's opinion, one of the great achievements was managing to complete the project scope a month before the final delivery, leaving time to improve the virtual environment presentation — a detail that wasn’t satisfactory in the initial versions.   In the project's bibliographic research, the group identified three tools for development: CrewAI and Langchain (for creating and managing AI agents) and the Unity game engine (for building the three-dimensional environment). The simplified structure of CrewAI facilitated the project's initial stages but became limited as computational memory needs increased. LangChain, in turn, allowed greater customization capacity in creating the agent. To no longer rely on a local database for testing, AWS Aurora's cloud was used.   After the initial research and discoveries, João Santos and Caio Bôa could focus on the program code, and Gustavo Oliveira and Matheus Castellucci, on the project documentation. They adopted the agile methodology, a series of values and project management principles born in the technology sector, and held monthly meetings with the advisor and bi-weekly meetings with mentor Rafael Perin from NTT DATA. WhatsApp, Discord, Microsoft Teams, and GitHub tools were used to maintain communication and organize task fulfillment.   Oliveira recalls that before the Capstone, the group’s knowledge about artificial intelligence was minimal. “I saw that various job offers in computing were related to AI,” he says, who now works as a junior data analyst in a fintech company. “In my job interview, I could show that I had the ability to learn a new tool in a short period and, from that knowledge, to generate a functional AI-focused product. And Insper's Capstone had a great influence on that.”   The learning is also contributing to Santos’s performance in the internship. “Here at the company, we are developing a project very similar to certain aspects of our Capstone,” he states. “I worked so much on that server structure we developed that it seems it became part of me. In the internship, I am replicating this structure, and it works very well. We had a huge progress in weeks.”   Castellucci, who ended up securing an internship at the same company as Santos, recalls that simply mentioning the Capstone topic already piqued the curiosity of interviewers in selection processes. The testimony aligns with the “mark on the resume” comment made by Dias from NTT DATA. “The more I talked about the project, the more people wanted to listen and paid attention, even if they didn’t fully understand what it was about,” Castellucci recollects.   Bôa had the same impression as his colleague during his internship interviews, even with companies not linked to information technology. A memorable moment for him was realizing the developmental possibilities of generative AI agents. “We finished the project scope early, but I think everyone can agree that the project itself is somewhat endless,” Bôa states. “If we were to think of more and more improvements, we wouldn’t have finished until now. This is a new technology, little used, and has much potential. We are satisfied with what we did.”"}]