TwiNSYNC

A context-aware digital twin system for personalized interview preparation.

TwiNSYNC is a context-enhanced digital twin system that simulates how you would answer interview questions — using your experience, tone, and content pulled in real time from multiple sources. It's powered by Model Context Protocol (MCP) principles to dynamically fetch and inject relevant context into every response.

What is TwiNSYNC?

TwiNSYNC assists in interview preparation by creating an intelligent representation of yourself. It functions as a digital model that understands your experience, communicates in your voice, and reflects your tone and style.

This system learns from your previous work, writing, and projects. When presented with a question, it generates answers that align with how you would respond — clear, confident, and personalized.

It integrates real-time information (such as your LinkedIn posts or YouTube presentations) with your resume and history to help you respond to challenging questions naturally and authentically, effectively coaching you through the interview process.

System Overview

User Query
   ↓
[1] Query Understanding → classify + extract topic
   ↓
[2] Prompt Blueprint Selection → build template
   ↓
[3] Context Fetchers (MCPs):
      - LinkedIn
      - X (Twitter)
      - YouTube
      - Memory DB (projects, past Q&A)
   ↓
[4] Prompt Composer → inject everything into prompt template
   ↓
[5] LLM or Agentic Framework → generate Twin's response
   ↓
[6] Output Display + Trace (sources used, feedback, rating)

Tech Stack

Component	Tech
Backend	FastAPI
LLM	OpenAI GPT-4o (can be swapped)
Memory	Chroma / SQLite + Sentence Transformers
Web Scrapers	Playwright / SerpAPI / Unofficial APIs
Optional UI	Streamlit, Gradio, or CLI
Agent Layer	LangGraph (optional)

Project Structure

twinsync/
├── app/
│   ├── main.py                # FastAPI entrypoint
│   ├── core/
│   │   ├── prompt_builder.py  # Injects context into prompt
│   │   ├── query_classifier.py# Extracts type & topic
│   │   └── twin_persona.yaml  # Stores your identity template
│   ├── mcp/
│   │   ├── linkedin.py        # Fetches LinkedIn data
│   │   ├── x.py               # Fetches Twitter posts
│   │   ├── youtube.py         # Gets transcripts
│   ├── models/
│   │   ├── memory_db.py       # Local memory (past answers, projects)
│   │   └── embeddings.py
│   └── routes/
│       └── ask.py             # Endpoint: /ask

Build Plan (MVP Phases)

Phase 1: Core Prompt System

• Define twin_persona.yaml (your static identity)
• Build one prompt template (e.g. STAR-based behavioral)
• Add basic query classifier (technical, behavioral, etc.)
• Create ask endpoint with FastAPI

Phase 2: MCP Context Fetchers

• Build mcp.linkedin.py: scrape/post fetcher
• Build mcp.x.py: grab latest tweets from your account
• Build mcp.youtube.py: transcript fetch from known talks
• Integrate memory_db.py (simple JSON or Chroma)

Phase 3: Prompt Composer

• Assemble dynamic context based on intent + query
• Inject all context into selected prompt template
• Final call to LLM + return formatted response

Phase 4: Output + Evaluation

• Show Twin’s response
• Include trace: source list, memory used, context injected
• Add optional evaluator to rate the twin's answer

Stretch Goals

• UI via Streamlit or Gradio
• Feedback loop: train twin via corrected answers
• Use LangGraph/CrewAI for advanced behavior simulation

Example Prompt Template

persona:
  name: "Zubin"
  role: "Product-Minded Technologist"
  tone: "Clear, warm, confident"

prompt_template: |
  You are {{persona.name}}, a {{persona.role}}.
  Respond with {{persona.tone}} in a job interview scenario.

  Role Context: {{role_applied}}
  Question Topic: {{interview_topic}}

  From Resume: {{memory_context}}
  LinkedIn Insight: {{linkedin_snippet}}
  From X: {{x_snippet}}
  YouTube Snippet: {{youtube_snippet}}

  Question: {{user_query}}

  Answer clearly and in your own voice.

Credits & Naming

Project Name: TwiNSYNC™

A context-aware MCP system that keeps your digital twin in sync with everything you are.

Next Step

To start the server:

uvicorn app.main:app --reload

To test with a mock question:

curl -X POST http://localhost:8000/ask -d '{"query": "Tell me about a time you led a design team."}'