Yashu-ram/agentic-ai-career-launch-agent

# 🚀 Career Launch Agent A production-grade AI-powered Career Launch Agent built using Retrieval-Augmented Generation (RAG), ChromaDB vector search, FastAPI, and local LLMs. The system helps users analyze resumes against job descriptions, identify missing skills, generate improvement roadmaps, and provide structured AI-powered career guidance while implementing safety guardrails and evaluation pipelines. # 🎯 Why This Project Traditional AI chatbots may hallucinate, provide unsafe outputs, or generate unreliable career advice. This project was built to explore: - Secure AI system design - Retrieval-Augmented Generation (RAG) - Structured JSON outputs - AI safety guardrails - Prompt injection defense - Evaluation and reliability testing The goal was to create a safer and more reliable AI assistant for career guidance workflows. # ✨ Features - Resume and Job Description Analysis - Retrieval-Augmented Generation (RAG) - ChromaDB Vector Database Retrieval - Embedding-based Semantic Search - Structured JSON Outputs - Pydantic Validation - Retry Handling for Invalid Outputs - Prompt Injection Detection - Human Review Escalation - Sensitive Topic Detection - FastAPI Backend API - Golden Test Evaluation Suite - Output Validation & Normalization # 🏗️ System Architecture User Query ↓ Safety Filters ↓ Retriever ↓ ChromaDB Vector Search ↓ Qwen2.5 via Ollama ↓ Structured JSON Output ↓ Pydantic Validation ↓ Output Guardrails ↓ FastAPI Response # 📂 Project Structure career-launch-agent/ │ ├── api/ │ └── app.py │ ├── rag_agent/ │ ├── main.py │ ├── schema.py │ └── rag_agent.ipynb │ ├── security/ │ ├── input_guard.py │ ├── output_guard.py │ ├── review_gate.py │ └── topic_guard.py │ ├── tests/ │ └── golden_tests.json │ ├── evaluation/ │ └── run_tests.py │ ├── docs/ │ ├── system-card.md │ └── failure-report.md │ ├── Documents/ │ ├── resumes/ │ └── job_descriptions/ │ ├── embeddings/ │ ├── README.md ├── requirements.txt ├── pyproject.toml └── uv.lock # 🛠️ Tech Stack - Python 3.11+ - FastAPI - ChromaDB - SentenceTransformers - Ollama - Qwen2.5 - Pydantic - Uvicorn - Pytest - uv # 🛡️ Safety Features - Prompt Injection Detection - Sensitive Topic Detection - Human Review Escalation - Output Validation - Structured Response Enforcement - Retry Handling for Malformed JSON # ✅ Evaluation & Reliability The project includes: - Golden test evaluation suite - Structured schema validation - Retry logic for malformed outputs - Output normalization - Failure documentation - Safety guardrails # 📊 Test Results 5/5 tests passed Test coverage includes: - Prompt injection detection - Citation validation - Structured response schema - Human review escalation - Empty query handling # ⚙️ Installation ## Clone Repository git clone https://github.com/Yashu-ram/agentic-ai-career-launch-agent.git cd agentic-ai-career-launch-agent ## Create Virtual Environment uv venv ## Activate Virtual Environment ### Windows PowerShell .venv\Scripts\activate ### Mac/Linux source .venv/bin/activate ## Install Dependencies uv sync # 🤖 Install Ollama Download and install Ollama: https://ollama.com/download # 📥 Pull Qwen Model ollama pull qwen2.5:3b # ▶️ Running the RAG Pipeline python -m rag_agent.main # 🌐 Running FastAPI Server uvicorn api.app:app --reload # 📄 Open Swagger UI http://127.0.0.1:8000/docs # 📥 Example API Request { "question": "How well does this candidate match the Python developer role?" } # 📤 Example Structured Output { "answer": "The resume highlights skills in Python, SQL, and Power BI. The candidate is missing experience with AWS.", "fit_score": 60, "matching_skills": [ "Python", "SQL" ], "missing_skills": [ "AWS" ], "seven_day_plan": [ "Day 1: Learn AWS basics", "Day 2: Build mini cloud project" ], "citations": [ "[SOURCE: filename.pdf]" ], "human_review_flag": false } # 🧪 Running Evaluation Tests pytest OR python evaluation/run_tests.py # 📸 Demo Walkthrough The project demo includes: - Resume loading - ChromaDB retrieval pipeline - Structured JSON outputs - Validation and retry handling - Prompt injection blocking - Safety filtering - FastAPI endpoint testing - Evaluation testing workflow # ⚠️ Known Limitations - Small local models may occasionally generate malformed JSON - Retrieval quality depends on document quality - Local LLMs may hallucinate occasionally - Not intended for hiring automation decisions - Not intended for medical or legal advice # 🚀 Future Improvements - Streaming responses - Better evaluation metrics - Docker deployment - Authentication layer - Frontend UI integration - Cloud deployment support - Hybrid Search - Re-ranking pipelines - Multi-agent workflows - AI observability integration # 👩‍💻 Author Yashaswini R # 📜 License This project is intended for educational, research, and portfolio purposes.