CareerAI: Personalized Career Trajectory Planning via Reinforcement Learning, Heterogeneous Graph Neural Networks and an Agentic LLM

Career guidance platforms currently treat professional development as a one shot retrieval problem: a user profile is ranked against available vacancies at a single point in time. This framing fails to model the sequential path dependent nature of career development in which each decision e.g., learning a skill, transitioning to a new role, earning a certification determines the feasibility of all subsequent decisions. This paper presents CareerAI, a system that reframes career guidance as a finite horizon Markov Decision Process (MDP) and solves it with a feasibility constrained reinforcement learning agent (MaskablePPO). A central contribution is the demonstration that goal achievement improves from 0% to 22%, and mean episodic return increases from +9.16 to +21.0, solely by replacing random role embeddings in the agent’s observation space with 64 dimensional embeddings produced by a heterogeneous GraphSAGE network trained on a career knowledge graph with no other changes to training. The system further integrates a three component hybrid recommender (Sentence-BERT content based, KNN collaborative and rule based expert components) with adaptive ensemble fusion. A ReAct pattern agentic LLM (Groq Llama 3.3 70B) exposing seven function calling tools and a perturbation based explainability layer providing SHAP inspired feature attributions, counterfactual skill analysis and prioritised learning roadmaps. GNN link prediction achieves AUC = 0.856 and Hits@10 = 0.733 on role transition edges. To our knowledge, CareerAI is among the first systems to unify GNN informed reinforcement learning planning, hybrid recommendation and grounded agentic LLM guidance with perturbation based explainability for personalised career trajectory optimisation.