Overview: Duck Debug is a web app that explains coding errors in plain language. Paste broken code and the error message, and the duck answers in a fixed, beginner-friendly shape: one sentence saying what went wrong, the line that caused it, why it happens, one guiding question, and one fix to try.
Problem statement: Error messages are written for experts. A beginner hitting their first traceback cannot tell which line failed, what the jargon means, or what to try next. Most tools hand over a copy-paste fix, which teaches nothing, so the same error comes back.
Proposed solution: A rubber-duck debugger powered by generative AI. The duck persona prompt forces plain language, zero unexplained jargon, answers under 200 words, and a warm, encouraging tone. Every answer teaches instead of just fixing, because the guiding question makes the learner think.
Technology stack: Node.js + Express 5 server, vanilla JavaScript frontend, Google Gemini API (gemini-3.1-flash-lite, free tier). No paid services. Playwright for screenshots and the demo clip. Vercel for hosting.
Implementation and demonstration: Live demo (link above once deployed) plus the demo video. The app has a no-key setup card, friendly error mapping for missing/bad keys, rate limits and timeouts, input validation, and XSS-safe rendering. New for DATANOVA: Error Insights. Every session logs only metadata (language, error signature, outcome, latency, whether the reply named the line); pasted code is never stored. /insights.html aggregates live sessions and the evaluation set. A sample-answer mode lets anyone try the duck with no API key, using a real recorded answer labeled as a sample.
Results and validation: Evaluation run Nov 10, 2026: 8 real debug sessions (Python 4, JavaScript 2, Java 1, C++ 1) against the Gemini API with the duck's exact prompt. 8 of 8 answered OK, 8 of 8 replies named the faulty line (100%), median answer time 2.3 seconds. Full dataset with the model's real replies: data/eval-sessions.json in the repo, produced by scripts/run-eval.py.
Potential impact: Beginners, students, and bootcamps get a patient explainer that is always available. Teachers can use the Insights page to see which errors a class hits most. Next steps: more languages, a 100-session evaluation, and a classroom mode.
Team contributions: Velu Murugan, solo: design, code, evaluation, video, PPT.
Pre-existing work, disclosed: Duck Debug was built Oct 17, 2026 for the Devpost "Build With AI: Basics" hackathon. The Error Insights logging, aggregation page, sample-answer mode, and the Nov 10 evaluation are new work done Nov 10, 2026 for DATANOVA.
Repo: https://github.com/Velu2k03/duck-debug — main commit ca7afd6950fcdd441be6082ddaa1f9f736fa15ba (Nov 10, 2026 evaluation run).
