Fengge Roleplay is an independent AI character chat product I took from the initial idea through product design, engineering, launch, and ongoing operation. It now serves thousands of registered users and has carried more than 20,000 messages in production, making it a living product rather than a staged demo.
Its source material is built from Fengge's video and livestream speech, rewritten into a curated question-and-answer corpus. An extraction pipeline normalizes the datasets, filters invalid entries, and deduplicates repeated questions while retaining the richer answer. The resulting 31,589 documents are embedded with Workers AI and indexed in Vectorize, with their full text and source metadata stored in D1.
At runtime, semantic retrieval and keyword scoring bring relevant material into DeepSeek's streaming responses. The experience also includes persistent conversation history, automatic or on-demand voice playback through MiniMax TTS, phone-based sign-in, and a focused Next.js interface deployed on Vercel.
Screenshot 1 of 3
