Hotmart uses classical machine learning where speed, cost, multilingual performance, or scale make an LLM a poor fit, while LLMs provide a more natural interface for customer-facing systems.
2
Hotmart Tutor answers questions from a course's own content and gives creators insight into what students are struggling with, while guardrails limit conversations outside the course.
3
AI platforms need independent components because models, vector databases, protocols, and evaluation methods change faster than a conventional ML platform can accommodate.
Summary
Paulo Vasconcellos describes Hotmart as a broad testing ground for applied AI. The company uses machine learning for fraud, content moderation, recommendations, forecasting, ticket classification, and entity extraction. It uses LLMs where users need a natural conversation, but keeps faster and cheaper classical models for high-volume tasks. Hotmart Tutor answers student questions from course content and shows creators which lessons need improvement. Hotmart also lets creators package their knowledge into agents they can sell. A sales agent on WhatsApp helps recover abandoned purchases, where response time matters. Paulo explains that these systems need direct user feedback, repeat engagement, retention, and business outcomes rather than vague claims about agent quality. He is candid about the infrastructure cost: AI platforms need retrieval, tools, fallbacks, vendor isolation, and components that can be replaced. The practical lesson is to choose technology for the problem, then expect to change the surrounding system as the field moves.
Hotmart applies machine learning to problems across the creator economy
Paulo describes Hotmart as a platform for online courses, tickets, ebooks, and other digital products. Its data teams work on fraud detection, spam detection, recommendations, forecasting, and moderation of harmful course content. He gives examples such as courses about buying firearms in Brazil or making drugs at home, which need to be detected and blocked. Hotmart's first production machine learning model went live around 2015, before the current generative AI wave.
Classical models remain the right choice for high-volume, narrow tasks
Paulo says Hotmart still uses spaCy and other classical approaches alongside LLMs. Entity extraction and ticket classification can run faster and at lower cost with models such as logistic regression, especially when thousands of customer tickets arrive each day. These models run as Kubernetes microservices without GPUs. They require ongoing training, evaluation, and checks for data drift and concept drift, but the operational effort can still be worth it. LLMs are preferred for the user-facing response because they provide a more human interaction.
Multilingual support makes small language models harder to use at Hotmart
Hotmart operates in Brazil and other regions, with most customer demand in Portuguese. Paulo says small language models can work well, but struggle more with multilingual use cases. This is one reason the company continues to use its established models. The choice depends on tradeoffs among quality, latency, infrastructure cost, training effort, and the volume of requests.
Hotmart Tutor combines course-grounded answers with creator feedback
Hotmart Tutor is a digital teacher that answers student questions about a course. It uses LLMs, retrieval-augmented generation, vector storage, fallbacks, and guardrails, and is restricted to the creator's content. The system gives students support when a creator cannot answer thousands of questions personally. It also analyzes the conversations for the creator. Repeated questions can point to a missing module, an unclear video, or a section that needs to be recorded again.
Guardrails matter when students take an agent outside its course context
Paulo describes students asking Tutor questions that are unrelated to the course, including personal or relationship advice. The agent needs strict guardrails so it can say that a request is outside its scope. He says most courses focus on subjects such as personal development, languages, or software development, but edge cases create risk. The agent should use relevant course context when it has it and avoid inventing a broader role for itself.
Hotmart measures agents through feedback, repeat use, and retention
Tutor includes thumbs-up and thumbs-down feedback for each answer. Hotmart also tracks whether people return and interact with the tutor during the week. Paulo argues that a large number of initial users does not prove value if poor answers stop them from returning. The company compares courses that use Tutor with courses that do not and watches student retention. He connects the interactive question-and-answer format with a shift from passive video consumption to more active learning.
Agent as a product lets creators sell knowledge through an interactive system
Hotmart launched an agent-as-a-product offering that lets creators build agents from their own knowledge bases and sell them alongside courses or ebooks. Paulo gives examples such as nutritionists creating agents that suggest recipes from course material. A creator can offer an agent as an additional product, including through WhatsApp. The aim is to let creators focus on making courses and sharing knowledge while agents handle some support and sales work.
AI platforms need replaceable components because the field changes quickly
Paulo says Hotmart had an established ML platform with feature stores, MLflow experiments, lifecycle management, and continuous training. Generative AI introduced new requirements such as vector databases, retrieval, model APIs, fallbacks, and agent tooling. He recommends keeping components independent so a vendor or storage system can be changed without rebuilding the whole platform. Hotmart moved from external vector database vendors to its own infrastructure by changing the endpoint behind the platform.
The right architecture depends on the actual constraint
Paulo argues that teams should isolate the problem before choosing a tool. A search tool may be enough in one system, while a private course knowledge base may need vector storage as well. Latency matters for a sales agent contacting someone through WhatsApp or assisting during a call, while a research answer can take longer if accuracy matters more. He warns against adopting a fashionable tool, such as a new data format or protocol, without a problem that justifies the work.
"Hotmart is a playground for data scientists because all the challenges opportunities that we see in data science, machine learning, we are tackling in Hotmart."Paulo Vasconcellos00:48
Who should watch
You run customer-facing AI systems and need to decide where an LLM adds value over a smaller or classical model.
You are building an agent platform and need practical guidance on retrieval, fallbacks, vendor changes, latency, and component boundaries.
You build education or creator products and want to connect user conversations with course improvement, support, sales, or retention.
