DevMentor AI

Inspiration

As students building projects on our own, we kept hitting the same wall: no mentor on hand at 1am when you're stuck on an idea, a bug, or how to explain your own code in an interview. We wanted something that acts like a patient, always-available senior developer — not a generic chatbot, but one scoped to the actual moments a student developer needs guidance: picking a project, planning it out, understanding code, fixing bugs, writing a README, and prepping for interviews.

What it does

DevMentor AI is a 6-in-1 mentor tool for developers, built around one shared AI backend:

Idea Generator — turns a rough idea into a scoped project brief with difficulty, time estimate, and tech stack
Roadmap — breaks a project into a day-by-day build plan
Code Explainer — explains what a code snippet does and why
Bug Fixer — diagnoses an error message and explains the fix
README Generator — turns project details into a polished, Markdown-formatted README
Interview Prep — asks a relevant technical interview question based on your project

How we built it

The frontend is React (Vite) with Tailwind CSS and React Router, talking to a FastAPI backend through one shared /api/generate endpoint. Every feature sends a mode + input and gets AI-generated text back — the backend picks the right system prompt per mode and calls the same single AI-calling function underneath. The backend runs on Google's Gemini API (via its OpenAI-compatible endpoint), deployed on Render; the frontend is deployed on Vercel.

Challenges we ran into

Almost everything that could teach us something did. A few real ones: our first API key nearly leaked to a public GitHub repo — GitHub's push protection caught it before it went live, which sent us down a real lesson in .gitignore and secret hygiene. We hit a wall of OpenAI billing requirements mid-build and pivoted to a free-tier provider instead, which meant restructuring our AI client without touching any other part of the app. We also learned the hard way that AI model names are a moving target — we had to build a small diagnostic script that tests several model names directly against our account to find one that actually still worked, after hitting multiple "model no longer available" errors in a row. On top of that: Windows terminal quirks (PowerShell vs Git Bash syntax differences), a Tailwind CSS major-version change that silently broke our styling for a while, and a very confusing detour where genuine backend crashes were being misreported by the browser as CORS errors.

Accomplishments that we're proud of

We shipped a fully working, fully deployed, 6-feature full-stack AI application — live on the real internet, not just running locally. More than that: we went from not knowing the difference between a terminal and a code editor to debugging real production issues (crash tracebacks, secret scanning, API rate limits, deployment configs) the way working developers actually do.

What we learned

Real debugging means reading the actual error message in front of you instead of guessing — the fix is almost always sitting in a traceback or a terminal, not in assumption. We also learned that "it's not working" always has a specific, findable reason, and that shipping software means constantly adapting around things outside your control (retired models, billing walls, platform quirks) without losing the core idea.

What's next for DevMentor AI

Wiring up shared project context so a student's idea auto-fills across all 6 pages instead of retyping it each time, adding accounts so people can save their generated roadmaps and READMEs, and upgrading to a paid-tier AI model for higher quality output once budget allows.