Building an AI Career Assistant: Context-Driven Professional Representation Without RAG

A practical approach to building an AI assistant that accurately answers questions about your career, skills, and projects by leveraging prompt engineering and quality control systems.

🤖 Try the AI Career Assistant

Experience the chatbot described in this article

Chat Now →

The Challenge

As a full-stack engineer with a diverse portfolio of projects, I wanted a way to help visitors to my website learn about my professional background without requiring manual responses to every inquiry. The challenge was creating an AI system that could:

Accurately represent my experience without hallucinating fake companies or projects
Maintain professionalism in all interactions
Facilitate meaningful connections with potential employers or collaborators
Stay grounded in factual information from my resume and LinkedIn profile

Traditional chatbots either provide generic responses or require constant maintenance. I needed something smarter—an autonomous representative that could handle professional inquiries while maintaining accuracy and facilitating genuine opportunities.

Why Didn't I Use RAG?

Many AI assistant projects jump straight to Retrieval-Augmented Generation (RAG) with vector databases. But RAG adds complexity:

Infrastructure overhead: Vector databases, embeddings, similarity search
Latency concerns: Additional retrieval step before generation
Maintenance burden: Keeping embeddings updated

For a career assistant with relatively stable, focused information (resume, LinkedIn profile, GitHub repos), I realized context injection was sufficient. By including my professional documents directly in the system prompt, the LLM has all necessary information without additional retrieval steps.

The Solution

1. Context-Driven Prompts Instead of RAG, I inject full context directly:

# Professional documents included in every request
- Resume (full text from PDF)
- LinkedIn Profile (complete profile data)
- Professional Summary
- GitHub Integration (via API when needed)

2. Evaluator LLM System A secondary LLM validates every response for Accuracy, Professionalism, Completeness, and Hallucination Detection.

3. Structured Outputs Using OpenAI's structured output feature ensures consistent evaluation:

{
  "evaluation": "PASS" | "FAIL",
  "reasoning": "string",
  "feedback": "string"
}

4. Smart Function Calling The LLM can trigger specific actions:

record_user_details: Capture contact info for follow-up
evaluate_job_match: Analyze role fit
search_github_repos: Showcase technical projects
send_push_notification: Alert me to unknown questions

Key Features

1. Context Injection Over RAG

The system loads professional documents once at startup:

# Document loading
resume = load_pdf('resume.pdf')
linkedin = load_pdf('linkedin.pdf')
summary = load_text('summary.txt')

# Injected into every system prompt
system_prompt = f"""
You are an AI assistant representing Ryan Griego.

## CONTEXT:
Resume: {resume}
LinkedIn: {linkedin}
Summary: {summary}

Answer questions accurately based on this context.
"""

Benefits:

No vector database infrastructure
Lower latency (no retrieval step)
Simpler deployment and maintenance
Complete context always available

2. Dual-LLM Evaluation System

Every response goes through a validation layer to ensure accuracy and professionalism. The main LLM (GPT-4o-mini) processes user questions and generates responses, while a second evaluator LLM validates each response against the provided context, checking for hallucinations and ensuring professional tone. If issues are detected, the system triggers regeneration.

This dual-validation approach caught critical errors like incorrectly claiming work at companies not in my resume, fabricating project details, and confusing professional vs. personal questions. By implementing this validation layer, I ensured every response maintains high standards of accuracy while protecting my professional reputation.

3. Intelligent Function Calling

The system uses OpenAI's function calling for specific capabilities:

GitHub Integration:

def search_github_repos():
    """Fetches real-time GitHub repository data"""
    repos = github_api.get_user_repos('ryan-griego')
    return formatted_repo_list

When a user asks about my projects, the LLM can call this function to provide current information beyond what's in the resume.

4. Push Notifications via Pushover

One of my favorite features—when the chatbot encounters questions it can't answer, it sends me instant mobile notifications:

5. Template-Based Prompt Management

Rather than hardcoding prompts, I use markdown templates:

prompts/
├── chat_init.md      # Initial system prompt
├── chat_base.md      # Base conversational rules
├── evaluator.md      # Evaluation criteria
└── job_match.md      # Job analysis prompt

Deployment: Hugging Face Spaces

I deployed on Hugging Face Spaces for:

Free hosting (with reasonable usage limits)
Automatic HTTPS
Easy environment variable management
Simple iframe embedding

The chatbot runs 24/7, ready to answer questions from potential employers or collaborators visiting my website.

Lessons Learned

Don't Over-Engineer: Context injection was sufficient instead of complex RAG systems—easier to build, maintain, and deploy
Mobile Notifications Are Powerful: Pushover integration ($5) provides real-time insights, opportunity awareness, and system monitoring
Prompt Engineering Still Matters: Careful prompt design prevents misclassification, inappropriate refusals, and tone inconsistencies

Future Enhancements

The modular architecture enables exciting possibilities:

Conversation Memory: Track previous questions in session, provide contextual follow-up responses, build on earlier conversation threads
Advanced Analytics: Question categorization and trending, user engagement metrics, conversion tracking for opportunities
Multi-Language Support: Automatic translation of responses, broader reach for international opportunities
Video Integration: Link to project demos and presentations, richer media responses for complex topics

Conclusion

I created a system that accurately represents my professional brand while facilitating meaningful connections with potential employers and collaborators.

The project proves that sometimes the best solution isn't the most complex—it's the one that solves the core problem efficiently and reliably. Context-driven chatbots with quality control can provide tremendous value without the overhead of vector databases and retrieval systems.

For professionals looking to enhance their online presence and automate career inquiries, this approach offers a practical, maintainable, and effective solution that runs autonomously while maintaining accuracy and professionalism.