CareerCompass AI: Multi-Agent Guidance & Skill Verification
Autonomous multi-agent platform orchestrating AI counseling, live technical challenge verification, and tailored career roadmaps.
Interactive Client-Side Demonstration • Features multi-agent dialogue states and interactive verification challenges.
What real-world challenge does this system address?
System Design & Approach
- Multi-Agent Separation of Concerns: Counselor, Verifier, Analyzer, and Roadmap agents collaborate with distinct system prompts and schemas.
- Interactive Skill Verification: Real-time code execution challenges and scenario questions that score candidate competency.
- Dynamic Milestone Generation: Roadmap synthesizer creating structured week-by-week technical goals.
- Persistent History: MongoDB Atlas storage maintaining multi-turn dialogue context and challenge progress.
System Architecture & Dataflow
User Client <-> FastAPI Router <-> Multi-Agent Dispatcher (Counselor, Verifier, Analyzer, Roadmap) <-> Gemini API & MongoDB Atlas
Streams user messages and renders dynamic challenge widgets and roadmap timelines.
Validates payload contracts, manages session tokens, and dispatches requests to appropriate agent instances.
Passes structured context between Counselor (dialogue) and Verifier (coding challenges) based on user intent.
Executes structured prompt generation with Pydantic JSON schemas and persists interaction logs.
Stack Specification & Components
Python 3.11 & TypeScript
FastAPI Backend & Next.js / Angular Frontend
MongoDB Atlas (M0 Cloud Cluster)
Vercel Frontend & Dockerized Backend Container
Multi-Agent State Orchestration Preview
Interconnected agents: Counselor, Verifier, and Roadmap Generator
“Candidate expressed interest in Applied Machine Learning engineering. Passing session context to Verifier Agent for hands-on challenge validation.”
“Implement a function to compute cosine similarity between two 1D NumPy arrays without relying on high-level Scikit-learn wrappers.”
“Roadmap Generated: Week 1-2 (Advanced Tensor Math) → Week 3-4 (Distributed FastAPI Pipelines) → Capstone Deployment.”
Engineering Decisions
Key design choices, constraints, and solutions formulated during development
Multi-Agent Decomposition over Monolithic Prompting
Divided the system into four discrete agents instead of a single prompt trying to counsel, test, evaluate, and roadmap simultaneously.
Single massive prompts suffered from instruction drift and loose schema compliance. Specialized agents allow independent prompt tuning, strict validation schemas, and modular unit testing.
Strict JSON Schema Enforcement via Pydantic
Configured LLM response generation with strict Pydantic JSON schema constraints.
Prevents conversational hallucinations from breaking frontend parsing in the challenge and roadmap UI components.