Inflearn's Poka: AI spam filter cuts daily spam accounts 76%

The account traces a progression rather than a single launch. Inflearn first relied on GPT-4o mini and a banned-word list, which worked for known patterns but required manual updates for each new spam type. The team then moved to a machine learning filter chosen for its ability to learn from limited data, fast inference, and easy retraining; SGDClassifier runs on CPU without a GPU, and a serverless design keeps cold starts around 2.8 seconds.

The filter assigns posts to three stages — spam, suspicious, or normal — and sends Slack alerts for the latter two. When the model struggled with Thai, Mongolian, Chinese, and Japanese spam, the team added GPT prompts to judge multilingual content. Later, to catch disguised ads and fake testimonials, it adopted a Zero Shot With Auto Generate Prompt method from Karrot's tech meetup, using GPT-4o to extract category features into MongoDB and a weekly Jenkins job to update them.

Beyond cutting manual work, Poka notes that removing spam posts linking to low-quality sites also improved SEO metrics. The team plans to upgrade the prompt management back office so operations staff can review and edit spam characteristics themselves.