# J.O.H.A.N(Just One Hilarious Answer Now) > Source: > Published: 2026-10-04 17:04:56+00:00 I built J.O.H.A.N — an AI-powered comeback generator designed specifically for a close friend of mine who often deals with a bully but never quite knows what to say back in the heat of the moment. My friend is a massive cinephile who loves Malayalam cinema, so I built J.O.H.A.N to listen to what the bully says and generate a handful of witty, adaptive comebacks in both English and Malayalam (Manglish). The AI uses a curated Knowledge Base of iconic Malayalam movie punchlines, trending memes, and actor profiles (like Mohanlal and Fahadh Faasil) to deliver culturally relevant, funny, and savage comebacks. It can even read them out loud! I also established a hard rule in the system prompt and a secondary safety filter to ensure the AI never generates jokes that may cross the line. (Insert a quick video here showing you speaking into the mic and J.O.H.A.N generating a comeback in the style of a Malayalam movie star!) **Built for a friend who needed to fight back with words.** Powered by open-source AI. Runs on your laptop. Accessible from your phone. **Savage Reply** is an AI-powered comeback generator that helps you respond to bullies with wit, humor, and devastating one-liners — in **English** and **Malayalam**. Your friend gets bullied and doesn't know what to say back? Just record what the bully said (or type it), and the AI generates **multiple witty comebacks** in different styles. Pick your favorite, and optionally have the AI read it out loud. J.O.H.A.N is built entirely around open-source AI and designed to run completely locally, even on my modest 4GB RAM laptop: The Brain (LLM): I used Gemma 2 (2B parameter model), Google's open-weight model, running entirely offline via Ollama. The 2B model is incredibly fast and fits perfectly within my limited RAM. The Knowledge (RAG): I built a lightweight Retrieval-Augmented Generation (RAG) system using Python. It injects a context window with JSON files containing 30+ classic Malayalam movie dialogues, 20+ meme templates, and specific actor humor profiles so the model understands the cultural vibe. Voice I/O: I used the open-source SpeechRecognition library for Speech-to-Text and edge-tts for free, local-friendly Text-to-Speech (which supports Malayalam voices!). The UI: A responsive, mobile-first dark-theme web app allowing my friend to access the local server during an encounter. Open innovation was the only way this project could exist in this form: Absolute Privacy: Bullying is an incredibly sensitive topic. Because I used an open-weight model (Gemma 2 2B) via Ollama, all data processing happens locally on the laptop. No API logs, no cloud telemetry, no risk of my friend's vulnerable moments being stored on a server they don't control. Cultural Customization: Closed models often struggle with hyper-local internet culture and Manglish (Malayalam written in English). Because I had full control over the open model and the local RAG pipeline, I could fine-tune the context window with exactly the right cultural references that resonate with my friend. Hardware Accessibility: Closed AI assumes you have a massive server or want to pay API fees. Open innovation allowed me to select a highly optimized, open-weight 2B model that runs beautifully on a basic 4GB RAM machine. I am submitting this project for the following categories: Main Prize: Build for a Friend - Built to solve a very specific, real-world social problem for a close friend. Best Use of Gemma - The core logic and comeback generation engine is powered by Google's open-weight Gemma 2 (2B) model running locally via Ollama.