MKishoreDev/PretendAi

PretendAI has one of the more original premises in this batch: instead of asking whether AI can sound human, it asks whether humans can convincingly sound like AI. The README explains that concept well, and the code backs it up with an actual game loop rather than just a landing page. The FastAPI backend exposes a clean /api/start, /api/reply, /api/finish, /api/leaderboard, and /api/health flow; sessions are persisted in MongoDB; Groq calls use fallback models and retry/backoff behavior; and the judge evaluates the player across helpfulness, neutrality, clarity, structure, accuracy, AI-likeness, hallucination risk, and empathy balance. I also like that the modes are meaningfully different: classic, interview, chaos, and jailbreak each change the kind of “human user” the AI role-plays. As a user, I can imagine starting a timed session, responding like an assistant, then getting a scorecard and leaderboard placement, which makes the product feel game-like instead of just a prompt demo. The biggest improvement would be tightening the project’s consistency: the README says Flask in one place, while the implementation is FastAPI, so the docs should be updated. I’d also add automated tests for session lifecycle, leaderboard insertion, invalid sessions, empty replies, and judge parsing. Since LLM judging can vary, storing example transcripts and expected score ranges would help keep future changes honest.