The problem_
Career advice is fragmented — resume building, job search, interview prep, and skill mapping all live in different tools. I wanted a single system where specialized agents collaborate, each owning a slice of the workflow.
What I built_
- Role-based multi-agent architecture — each agent has a narrow scope (resume, search, interview, skills) and a shared memory layer.
- Dynamic routing — LangGraph coordinates handoffs based on conversation state and user intent.
- Short-term memory for contextual decisions across turns; long-term memory for user profile + history.
- Intelligent routing heuristics to avoid agent overlap and reduce token cost.
Stack_
- LangGraph for orchestration
- Qwen3 as the reasoning model
- Python backend, lightweight React front-end
- Vector store for long-term memory
Lessons_
Multi-agent systems live or die by their routing layer. Good prompts alone aren’t enough — explicit state machines and guardrails beat cleverness every time. Memory design was the biggest win: scoping memory per agent kept answers crisp.
─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─