A Junior DevOps Engineer's Experience Introducing AI/ML-Based Automated Spam Filtering

Inflearn introduced an AI/ML-based automated system instead of manual review to solve the problem of surging spam posts as the platform grew. Initially, they approached this with a Few-Shot method using GPT-4o-mini and a banned word list, but felt limitations in responding to new patterns and handling multiple languages.

To improve this, they vectorized text with TfidfVectorizer and built an adaptive spam filter using SGDClassifier. This model has low computational requirements, showing a very fast inference speed of 0.2~0.6ms in the AWS Lambda environment, and was designed to enable continuous learning through S3 and Jenkins.

Additionally, to respond to multilingual spam, they combined it with ChatGPT to increase discrimination accuracy, and introduced a 3-tier classification system of 'spam', 'suspicious', and 'normal'. Posts in the 'suspicious' tier minimized the operations team's management workload through Slack notifications and emoji automation.