PathFinder AI

"Your future, mapped. One smart step at a time."

Comment

πŸ“– About the Project

Inspiration

The modern student faces an overwhelming paradox: there are over 26,000 universities worldwide and yet most applicants apply to the wrong schools β€” either reaching too far or settling too low β€” simply because they lack the right data to guide them.

Beyond university selection, many young people graduate without a clear vision of their career trajectory. They accumulate degrees and grades but struggle to translate their profile into a meaningful professional path.

This project was inspired by the intersection of three observations:

  1. Information asymmetry: High-achieving students in developing countries often have no access to credible, comparative university data.
  2. The "fit" problem: Generic ranking systems (QS, THE) rank universities globally, but ignore the student's individual profile. A student with a 3.2 GPA, passion for robotics, and a hackathon award needs different advice from one with a 3.9 GPA targeting medical school.
  3. Career blindness: There is very little tooling that bridges the gap between a student's academic profile and real-world career outcomes.

"Every student deserves a guide who knows both who they are and where the world is going."

PathFinder AI aims to be that guide β€” powered by data, personalized by AI.

What it does

PathFinder AI is an intelligent web platform that serves two core functions:

  • University Recommendation: Students input their GPA, skills, extracurricular achievements, preferred country, and budget. The system computes a compatibility score against a database of 5,000+ global universities and returns a ranked list of best-fit institutions β€” complete with acceptance likelihood, tuition info, scholarship opportunities, and program highlights.
  • Career Roadmap Generation: Based on the student's profile and career interests, the system generates a personalized, step-by-step professional development plan β€” including skills to acquire, online courses to take (Coursera, edX), certifications to earn, and projected salary growth over time.
  • AI Chat Assistant: A GPT-4o powered chatbot allows students to ask natural language questions such as "How do I improve my chances for a top-10 CS school?" or "What career suits someone with my background in biology and data science?"
  • Skill Gap Analysis: Visual breakdown of what skills a student currently has vs. what their target universities or career paths require.

How we built it

PathFinder AI is built as a microservices-based monorepo across three main layers:

Frontend β€” React.js (Vite) with TailwindCSS and Framer Motion for a fast, animated, and fully responsive UI. Chart.js and Recharts power the skill radar charts and salary projection graphs. KaTeX renders math formulas inline.

Backend API β€” Node.js + Express.js serves as the API gateway handling authentication (JWT), rate limiting, and request routing. Prisma ORM connects to SQL Server (local) as the primary relational database, while MongoDB stores unstructured AI-generated career logs.

AI/ML Microservice β€” A separate Python FastAPI service handles the core intelligence:

  • GPA Matching: Normalized distance scoring between student GPA and university admission averages.
  • Skill Matching: TF-IDF vectorization + Cosine Similarity between student skill sets and university program requirements.
  • Achievement Scoring: Weighted scoring across achievement categories (research papers, hackathons, sports, etc.).
  • Career Roadmap: GPT-4o API with structured prompt engineering generates the personalized narrative roadmap.
  • Collaborative Filtering: scikit-learn models trained on historical admission data refine recommendations over time.

Redis is used to cache recommendation results and reduce redundant ML inference and GPT API calls.

Challenges we ran into

  • Data Quality & Availability: University admission data (average GPA, acceptance rates) is inconsistently published, especially for non-western institutions. We addressed this by combining QS/THE API data with web scraping and community-contributed records.
  • Cold Start Problem: New users with minimal profiles yield poor collaborative filtering results. We implemented a hybrid approach β€” content-based matching (GPA + skills) for new users, gradually shifting to collaborative filtering as the user base grows.
  • GPT-4o Cost at Scale: Running GPT-4o for every roadmap request is expensive. We mitigated this with Redis caching (keyed by profile hash), per-user rate limiting, and a Retrieval-Augmented Generation (RAG) approach to reduce prompt token size.
  • Cultural & Geographic Bias: Initial training data skewed toward Western universities. We explicitly sourced datasets from Asian, African, and Latin American institutions and integrated regional ranking systems (ARWU, Webometrics).
  • Privacy & Compliance: Academic profiles are sensitive data. We implemented AES-256 encryption at rest, TLS 1.3 in transit, and GDPR-compliant data deletion workflows.

Accomplishments that we're proud of

  • 🎯 Built a composite recommendation algorithm combining GPA matching, cosine-similarity skill matching, and achievement scoring into a single weighted score β€” all explainable to the end user.
  • πŸ—ΊοΈ Delivered AI-generated career roadmaps that are genuinely personalized, not just generic career advice β€” using structured GPT-4o prompts that incorporate the user's exact profile.
  • πŸ“Š Created a skill gap visualization that visually shows students exactly what they need to improve to reach their target universities.
  • 🌍 Curated a database spanning 100+ countries to avoid the trap of only recommending US/UK institutions.
  • ⚑ Achieved fast recommendation responses via Redis caching, reducing repeated ML inference to near zero latency for returning profiles.

What we learned

  1. Recommendation systems are fundamentally a data problem β€” even the best algorithm fails without clean, unbiased data.
  2. User onboarding friction directly affects data quality β€” simpler registration (name + age first) proved critical to getting users to engage before asking for deeper profile information.
  3. AI narrative generation requires careful prompt engineering β€” generic GPT prompts yield generic career advice; structured, profile-aware prompts yield genuinely useful roadmaps.
  4. Microservices require discipline β€” separating the ML service (Python) from the main API (Node.js) adds operational complexity but dramatically improves scalability and maintainability.
  5. Visualizing data builds trust β€” users are far more likely to act on recommendations when they can see why a university was matched to them (score breakdown, skill gap chart).

What's next for PathFinder AI

  • 🌐 Multi-language support β€” Vietnamese, Chinese, Spanish interfaces to serve our global audience natively.
  • πŸ“± Progressive Web App (PWA) β€” Offline-capable mobile experience for students in low-connectivity regions.
  • 🀝 Mentorship Matching β€” Connect students with alumni from their target universities via a community feature.
  • πŸ“ˆ Live Job Market Data β€” Real-time job posting trends integrated from LinkedIn and Indeed APIs to keep career roadmaps up to date.
  • πŸŽ“ Scholarship Discovery Engine β€” Dedicated module to surface scholarship opportunities matched to the student's nationality, GPA, and program of study.
  • 🏫 Institutional Partnerships β€” Partner with universities to provide official recruiter-side dashboards for talent discovery.

Built With

Share this project:

Updates