AI Interview Companion (Second Brain)

Technologies: Cloudflare Workers, Durable Objects, Vectorize (RAG), Llama 3.3, TypeScript, Python

Description

• RAG Architecture: Built a serverless Retrieval-Augmented Generation system using Cloudflare Vectorize. It converts user notes (STAR stories, technical definitions) into 1024-dimensional embeddings, allowing the Llama 3.3 LLM to recall specific personal experiences during mock interviews.
• Engineering Rigor: Developed a custom Offline Evaluation Pipeline in Python (sentence-transformers). This automated benchmark regression-tests the agent’s memory, validating a 90.3% semantic accuracy against a “Golden Dataset” of ground-truth answers.
• State Management: Implemented Durable Objects to manage real-time WebSocket connections and chat history, ensuring consistent context retention across distributed edge locations.

Figure: Automated benchmarking results showing >90% accuracy in retrieving Personal History and ML concepts.