CareerForge-AI

AI-powered career copilot that turns your target role and skill profile into a concrete roadmap, portfolio plan, resume checklist, and weekly action steps.

Comment

Inspiration

Most students know what they want to become but have no idea where to start. Career advice online is fragmented — one blog says learn Python, another says build projects, another says network. There's no unified, personalized execution system. CareerForge AI was built to fix that: one input, one complete action plan.

What it does

CareerForge AI is a student career-prep copilot. You enter:

  • Your target role (e.g., Software Engineer, Data Analyst)
  • Your current skills, strengths, and weak areas
  • Your weekly time budget

The app generates a complete, role-specific package:

  • Skill-gap analysis — what's missing vs. what the role requires
  • Personalized learning roadmap — prioritized skills to acquire
  • Portfolio project ideas — concrete builds that signal competence
  • Resume improvement checklist — actionable fixes, not vague tips
  • Interview prep set — role-specific questions with focus areas
  • Weekly execution plan — a structured schedule you can start today
  • Progress tracker — persistent across sessions via localStorage ## How we built it | Layer | Technology | |---|---| | Frontend | Next.js 14, TypeScript, Tailwind CSS | | AI Backend | NVIDIA NIM API (LLM inference) | | API Layer | Next.js Server Route /api/generate | | Validation | Zod schema validation | | Deployment | Vercel |

The core flow: the frontend collects a ProfileInput object → sends it to /api/generate → the server calls NVIDIA NIM with a structured prompt → the JSON response is sanitized into a typed PlanOutput shape → rendered as dashboard cards.

To ensure judges could always test full functionality, I built a deterministic fallback mode: if the NVIDIA NIM API key is missing or returns an error, the app serves a structured mock plan instead of breaking.

Challenges we ran into

Prompt engineering for structured JSON output — getting the LLM to consistently return valid, complete JSON across all output sections required iterative prompt refinement and a post-processing sanitization layer.

  • Resilient error handling — the frontend needed to gracefully handle non-OK API responses, malformed payloads, and network failures with inline user-visible errors rather than silent crashes.
  • Demo reliability — ensuring judges could test the full flow without needing to configure API keys led to the fallback seed profile + mock output system. ## What we learned How to structure LLM output via prompt design and schema enforcement with Zod
  • Production-grade error boundaries for AI-powered Next.js apps
  • NVIDIA NIM API integration and inference reliability patterns
  • The value of demo-first design in hackathon builds

Built With

Share this project:

Updates