CodeBuddy

💡 𝗪𝗵𝘆 𝗪𝗲 𝗕𝘂𝗶𝗹𝘁 𝗜𝘁 — Project Story: As developers, we spend a massive chunk of our time debugging syntax errors, deciphering cryptic edge cases, and waiting for peers to review our pull requests — sometimes for days. Students face this even harder: no senior dev to ask, no reviewer on standby, just a blinking cursor and a confusing error message at 2 AM. We wanted to build a tool that acts as a 𝟐𝟒/𝟕 𝐝𝐢𝐠𝐢𝐭𝐚𝐥 𝐩𝐞𝐞𝐫 𝐫𝐞𝐯𝐢𝐞𝐰𝐞𝐫 — something that instantly highlights logical flaws, scores code quality, suggests optimizations, and lets you chat directly with your code. No waitlists. No pull requests. Just paste and learn.

🛠️ 𝗛𝗼𝘄 𝗪𝗲 𝗕𝘂𝗶𝗹𝘁 𝗜𝘁 𝗦𝘁𝗮𝗰𝗸: Python · Streamlit · CSS · Gemini 1.5 Flash · FPDF2 · python-dotenv The application uses a modular architecture that cleanly separates the static code analysis engine from the conversational chatbot interface. 🖥️ 𝗙𝗿𝗼𝗻𝘁𝗲𝗻𝗱 — Streamlit , CSS A dual-panel UI — a Review Panel where users paste code and trigger an instant structured audit, and a persistent Chatbot Panel that maintains conversation history for follow-up questions like "Can you rewrite that using a list comprehension?" 🧠 𝗔𝗜 𝗘𝗻𝗴𝗶𝗻𝗲 — Gemini API Powered by google-generativeai with engineered system prompts that force the model to behave like a strict but constructive Senior Software Engineer. Native Gemini chat session objects handle context memory without manually parsing token histories. 📄 𝗣𝗗𝗙 𝗥𝗲𝗽𝗼𝗿𝘁𝘀 — FPDF2 Generates downloadable review reports complete with score tables, issue breakdowns, AI-improved code, and personalized learning tips.

🚧 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲𝘀 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗚𝗲𝗺𝗶𝗻𝗶 𝗢𝘂𝘁𝗽𝘂𝘁 — Getting Gemini to return clean, parseable JSON consistently — instead of a wall of markdown prose — required careful prompt engineering and robust fallback parsing logic. 𝗦𝘁𝗿𝗲𝗮𝗺𝗹𝗶𝘁 𝗦𝘁𝗮𝘁𝗲 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 — Streamlit reruns the entire script on every user interaction. Mastering st.session_state was critical to ensure chat memory, scores, and code context persisted without wiping on every click. 𝗕𝗲𝗴𝗶𝗻𝗻𝗲𝗿-𝗙𝗿𝗶𝗲𝗻𝗱𝗹𝘆 𝗧𝗼𝗻𝗲 — Tuning the system prompt to explain issues in simple language — without being condescending or losing technical accuracy — required multiple iterations and real student test cases.

💡 𝗪𝗵𝗮𝘁 𝗪𝗲 𝗟𝗲𝗮𝗿𝗻𝗲𝗱 ① 𝗣𝗿𝗼𝗺𝗽𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝗼𝗿 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗢𝘂𝘁𝗽𝘂𝘁𝘀 Constraining LLM responses using explicit JSON schemas transforms chaotic AI output into clean, renderable UI data. The format of your prompt is just as important as the content. ② 𝗦𝘁𝗮𝘁𝗲 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 𝗶𝗻 𝗦𝘁𝗿𝗲𝗮𝗺𝗹𝗶𝘁 Properly scoping st.session_state keys is essential for any stateful app — chatbot memory, scores, and user code must all be explicitly preserved between reruns. ③ 𝗗𝗲𝘀𝗶𝗴𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗕𝗲𝗴𝗶𝗻𝗻𝗲𝗿𝘀, 𝗡𝗼𝘁 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿𝘀 Most dev tools assume prior knowledge. Building for learners forced us to rethink every label, error message, and feedback card — clarity always beats precision when your user is still learning. ④ 𝗠𝗼𝗱𝘂𝗹𝗮𝗿 𝗔𝗜 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗣𝗮𝘆𝘀 𝗢𝗳𝗳 Separating Gemini logic into gemini_helper.py made it trivial to swap models, adjust prompts, and debug AI responses without ever touching the UI layer.

🔭 𝗪𝗵𝗮𝘁'𝘀 𝗡𝗲𝘅𝘁 Monaco-based in-browser code editor · Teacher dashboard for batch student reviews · Gamified XP and badge system to encourage daily practice · GitHub integration for commit-level automated feedback