Amit Kumar
Palo

Smart Entertainment BOT is a ChatGPT-style movie, series & anime recommender I built end-to-end. "What should I watch?" is a hard question — you don't know the title, you know the mood. Describe it ("something unsettling, but not too gory") and it asks a couple of quick questions, finds real, catalog-verified titles, and tells you the story behind each pick so you can feel the film before spending two hours on it.

• Grounded with RAG, not fine-tuning — live catalog APIs (Watchmode, OMDb, AniList, Jikan, TVMaze). Every title is validated before it's shown, so it cannot hallucinate a film, rating, or cast — ask for something that doesn't exist and it says so.
• Two-tier LLM routing — OpenAI GPT-OSS (120B & 20B) on Groq; the large model only runs when the turn earns it, while tool calling drives search, clarifying questions, summaries, reviews, cast and "more like this".
• Cost engineering — three Redis cache layers (24h TTL) shared across users: the second person to ask what the first already asked costs zero tokens. Failed answers are never cached, so one bad turn can't poison a question.
• Guardrails — catalog-validated titles only, polite off-topic declines, and quality ranking (rating² × popularity) so "thriller" never returns a random crime drama.
• Vision extension (v2) — drag, paste or upload a movie poster and Google Gemini identifies the film. The guess is treated as a hypothesis, then catalog-verified before display — a wrong vision guess can't leak into the chat.
• Experience & observability — SSE token streaming with typing indicator, clarifying-question chips, resumable sessions with LLM-generated titles, returning-user recognition, and Langfuse tracing for prompts, tool calls, latency, tokens, cost & cache hit/miss per turn.