Vibe Bro

Paste the code your AI wrote and say what you asked for. Vibe Bro takes a second look.
Vibe Bro flags a fake package and a failing test: the CSV header is counted as an order.
No answer dump. You predict first, then get a hint, then see the faulty line, then the fix.
Fix the code and re-run the same tests. When they all pass, the fix is proven, not guessed.
Short drills adapt to the bug types you miss most. Pick the buggy line and name the bug type.

Inspiration

We built Vibe Bro with Claude as our AI coding assistant, and we saw first-hand how easy it is to ship code you don't fully understand. Students and freshers who use ChatGPT, Claude or Gemini to write code often can't tell when that code is wrong. We wanted a tool that checks the code and also helps them learn from the mistakes.

What it does

You paste code an AI wrote and say what you asked for. Vibe Bro then:

Checks that Python can read the code
Checks that every imported package exists on PyPI, which catches invented packages
Checks library calls against live docs to catch deprecated or removed functions
Writes tests from what you meant the code to do, not from what it happens to do
Runs those tests in a sandbox

For each problem, it asks you to predict what will happen, gives a hint, and hides the faulty line and the fix until you ask. You then fix the code yourself and re-run the same tests.

There is also a Practice drill with 14 short buggy snippets. Vibe Bro tracks which kinds of bugs you miss and shows you more of those.

How we built it

All AI calls go to NVIDIA Nemotron models through Nebius Token Factory.

Nemotron 3 Nano handles the many short jobs: reading docs and explaining bugs.
Nemotron 3 Super handles the harder jobs: writing tests and judging your explanations.
Tavily searches live documentation.
E2B runs pasted code in a disposable cloud sandbox on the hosted site. On a laptop it runs in a throwaway folder with time and memory limits.
Streamlit and Python make up the app.
Claude was our AI coding assistant. We used it to help write the app, which is why we built a tool that checks AI-written code.

Challenges we ran into

Running strangers' code safely. The hosted app uses a disposable E2B sandbox that never sees our API keys.
Not trusting the AI to judge correctness. AI writes the tests, but the computer decides pass or fail. A failing test is proof of a bug, not a guess.
Teaching instead of telling. We built the flow around hints first, answers later.

Accomplishments that we're proud of

Every bug Vibe Bro reports is backed by a failing test. We also built the whole loop, from the first check to the fix and re-run, plus a drill that adapts to what you get wrong.

What we learned

AI can be confidently wrong, so code has to be run to be trusted. Matching model size to the job (Nano for short tasks, Super for hard ones) keeps things fast and cheap.