Interview Prep Companion (cf_ai_second_brain)

A full-stack RAG (Retrieval Augmented Generation) agent built on Cloudflare Workers, Durable Objects, and Vectorize. It serves as a personalized "Interview Prep Companion" that helps students store and recall behavioral stories, algorithm patterns, and system design notes.

🚀 Live Demo

https://cf-ai-second-brain.rnazre.workers.dev

🛠️ Tech Stack

• LLM: Llama 3.3 70B (via Workers AI)
• Coordination: Durable Objects (managing chat state and agent logic)
• Memory: Cloudflare Vectorize (Vector Database) + BGE-M3 Embeddings
• Frontend: React + Vite (served via Worker Assets)
• Language: TypeScript
• Evaluation: Python (sentence-transformers, requests)

✨ Features

1. Mock Interviewer: Real-time streaming chat with Llama 3.3, customized to act as a supportive technical interviewer.
2. Study Guide Mode: Users can save specific notes (STAR method stories, LeetCode patterns) using the /learn command.
3. RAG / Memory: The agent converts /learn inputs into vector embeddings (1024 dimensions) and stores them in Cloudflare Vectorize. When asked a question later, it retrieves the user's specific saved examples rather than generic advice.
4. Persistent History: Chat sessions are managed by Durable Objects, ensuring state consistency.
5. Automated QA: Includes a Python-based evaluation suite to regression test the agent's memory.

🧠 Example Usage & Demo

The system is designed to retain complex technical details and recall them during mock interviews.

1. Saving Knowledge (/learn) The user stores a definition or behavioral story into the Vectorize database.

/learn Concept: Big O notation...

2. Recalling & Reasoning The agent retrieves the stored definition to answer complex reasoning questions.

User: "Which algorithm likely has better Big O complexity?"

🏃‍♂️ How to Run Locally

Prerequisites

• Node.js & npm
• Cloudflare Wrangler CLI (npm install -g wrangler)
• Python 3.8+ (for running benchmarks)

Installation

1. Clone the repository:

   git clone [https://github.com/rukmini-17/cf_ai_second_brain.git](https://github.com/rukmini-17/cf_ai_second_brain.git)
   cd cf_ai_second_brain

2. Install dependencies:

   npm install

3. Login to Cloudflare:

   npx wrangler login

4. Setup Vector Database:

   npx wrangler vectorize create second-brain-index --dimensions=1024 --metric=cosine

5. Run Development Server:

   npm run dev

📊 Evaluation & Benchmarking

To ensure reliability and prevent hallucinations, I implemented an automated offline benchmarking pipeline.

• Methodology: A "Golden Dataset" of 12 ground-truth Q&A pairs covering Behavioral, Technical (ML/Security), and Personal domains.
• Metric: Semantic Similarity (Cosine Similarity) measured using the all-MiniLM-L6-v2 embedding model.
• Performance: The system achieved an overall accuracy of 90.3%, with near-perfect retrieval for Resume facts (93.9%) and ML concepts (91.3%).

You can reproduce these results by running the script in the evaluation/ folder:

cd evaluation
pip install -r requirements.txt
python bulk_benchmark.py 

See benchmark_results.csv for the raw data.

📂 Project Structure

• src/server.ts: Main Worker and Durable Object logic (RAG implementation).
• src/client.tsx: React frontend.
• evaluation/: Python benchmarking scripts and datasets.
• wrangler.jsonc: Cloudflare configuration (Vectorize bindings, Assets).

🤖 AI Assistance

Prompts used during development are documented in PROMPTS.md.