Career Compass

Most career-advice tools either ask forty questions upfront or generate the same five bullet points regardless of who's asking. Career Compass tries to sit somewhere in between — a multi-agent system, orchestrated with CrewAI, where each agent owns a different part of the assessment: reading a resume, checking it against current market signals, evaluating skill gaps.

Grounding was the part I cared about most. Rather than let the agents free-associate career advice, I built a LangChain history-aware retriever over a FAISS vector store — documents hosted on S3, chunked at 1000 characters with 100-character overlap, embedded with Google's embedding-001 — so recommendations trace back to something real rather than something merely plausible-sounding. Generation runs on Llama 3.3 70B via Groq.

Session state persists per user across visits with Flask and SQLite, and cross-turn memory runs through CrewAI plus Mem0, so the system remembers what it already told you the last time you asked.