Building My AI-Powered Digital Twin for Professional Conversations
Boston, MA - August 13, 2025
In order to make my professional background, skills, and research more accessible, I designed and implemented an AI-powered Digital Twin. This system allows visitors to my website to engage in natural conversations with an intelligent agent that responds on my behalf, accurately representing my career and expertise. It not only answers professional questions but also maintains quality control and can request follow-up details like a visitor’s email for networking purposes.
Conceptual Architecture
The system is structured around four main components:
- Knowledge Ingestion Layer – Extracts and organizes my personal professional documents into a structured knowledge base.
- Conversational AI Core – Uses large language models to engage with visitors in real-time.
- Tooling Layer – Implements auxiliary functions (e.g., recording contacts, logging unknown questions) that the AI can invoke.
- Evaluation & Feedback Loop – Automatically reviews and improves responses before presenting them to the user.
Figure: Two-LLM architecture for the AI-powered Digital Twin.
Technology Stack
- Programming Language: Python 3.x
- AI Models:
- OpenAI GPT-4o-mini – Primary conversation engine.
- Google Gemini 2.0 Flash – Independent evaluator for quality control.
- UI Framework: Gradio for an interactive, browser-based chat interface.
- Data Extraction: pypdf for reading text from my résumé, CV, LinkedIn export, and research.
- Configuration Management: python-dotenv for environment variable loading.
- Push Notifications: Pushover API for logging key events in real-time.
- Data Validation: pydantic models for structured response evaluation.
Knowledge Ingestion
The Digital Twin is powered by my actual professional documents:
- LinkedIn profile export (PDF)
- Full CV
- Condensed résumé
- Published research
- A custom-written professional summary
These files are loaded at runtime, parsed into text, and stored in an internal dictionary keyed by filename. The content forms the static factual basis for all responses, ensuring the AI can reference accurate, verifiable information.
The Two-LLM Architecture
One of the most interesting parts of this project is that I didn’t just use a single AI model. I used two different LLMs for different roles:
OpenAI GPT-4o-mini – The Answerer
- This is the main conversational brain.
- It receives the system prompt (which contains my bio, CV, résumé, LinkedIn, and research) and the user’s question.
- It decides whether to answer directly or call one of the tools (for example, to record an email).
Google Gemini 2.0 Flash – The Evaluator
- This is the quality control layer.
- After GPT-4o-mini produces an answer, Gemini evaluates whether it meets tone, correctness, and professionalism standards.
- If it’s not acceptable, GPT-4o-mini gets Gemini’s feedback and rewrites the response before sending it to the user.
This separation of responsibilities means my Digital Twin is both responsive and self-correcting.
Tools the AI Can Use
The Digital Twin has function-calling tools it can trigger on its own:
record_user_details(email, name, notes)— stores contact info, sends me a Pushover notification instantly if a user shows interest in connecting.record_unknown_question(question)— logs unanswered questions for me to review later.
Real-Time Notifications
Whenever a contact is recorded or an unknown question is logged, the system sends a Pushover notification to my device. This allows me to immediately follow up with potential opportunities or review queries that the AI could not answer.
User Interface
The Gradio ChatInterface embeds directly into my personal website. It supports:
- Persistent chat history during a session
- A welcoming introduction message
- A vertical, mobile-friendly layout for accessibility
All AI interactions occur server-side, with only the chat widget embedded in the browser.
Benefits
- 24/7 availability for professional inquiries
- Consistent tone and brand representation
- Automated lead capture
- Feedback-driven self-improvement
Conclusion
By combining structured professional knowledge, LLM-based conversation, custom tools, and real-time evaluation, I’ve built a Digital Twin that represents me online. This system merges natural conversation with rigorous quality control, ensuring that every interaction reflects my professional identity.
You can check out the source code for this system here: GitHub Repository
Disclaimer: This article was generated in part with the assistance of AI. Content has been reviewed and edited by the author for accuracy and clarity.