What I Built
I built LocalMock, a completely offline, multi-agent technical interview practice partner for my friend and frequent project collaborator, Gayatri Akalwadi.
Gayatri and I often build AI tools together, and she is currently gearing up for intense Machine Learning Engineer interviews. She wanted a practice partner that could grill her on complex RAG architectures and MLOps pipelines. However, she had two major concerns: she didn't want to upload her personal resume and detailed performance history to a closed-source cloud model, and she didn't want to worry about racking up API usage costs during marathon 4-hour practice sessions.
LocalMock solves this by providing a conversational, rapid-fire mock interviewer that runs entirely on her laptop. When she handed over her resume, the system dynamically generated a custom interview loop, adjusting the difficulty based on her real-time answers—all without a single byte of data leaving her local machine. When I installed it for her, she said, "This is exactly what I needed. I can finally mess up and learn without feeling like a tech giant is logging my mistakes."
How I Built It
The core of LocalMock is entirely driven by open-source AI technologies:
Open-Weight Models: I used Ollama to serve open-weight models locally. The application uses Llama-3.2-3B for rapid, low-latency conversational replies and Gemma-2-9B for deep, end-of-session feedback and grading.
Agent Framework: I built the orchestration logic using the open-source LangGraph framework. It features a multi-agent setup: a "Technical Interviewer" agent that tracks the conversation state and poses questions, and a "Critique" agent that silently evaluates the accuracy of the user's answers.
Local RAG: I utilized ChromaDB (open-source vector database) and local embeddings to process Gayatri's resume and past project files, allowing the Interviewer agent to ask highly contextual questions (e.g., "I see you used Databricks here—how would you handle streaming data latency in that setup?").
Frontend: A lightweight React SPA that interfaces with the local Python backend.
Why Does Open Innovation Matter?
For this project, open innovation wasn't just a feature; it was the entire foundation of the solution:
Absolute Data Privacy: Resumes contain highly personal information (phone numbers, addresses, work histories). By using open-weight models and local inference, Gayatri's data stays 100% on her hardware. A closed API would have required trusting a third-party server with her personal life.
Zero Operating Cost: Running continuous back-and-forth multi-agent loops with a cloud provider would get expensive fast. With local models, she can practice for 10 hours a day for free.
Hardware Freedom: Open innovation meant I could configure the agents to run smoothly on the hardware she already owns, keeping the memory footprint manageable without relying on an internet connection.
Prize Categories
Open Source AI Vanguard
Local Inference Champion
Creative Use of Agent Frameworks
Top comments (0)